- The paper introduces ThermBuild, pairing 15-minute measurements from two instrumented German homes with 958 validated TRNSYS simulations covering diverse buildings, climates, HVAC systems, and occupant behaviors.
- The paper validates the simulations against measurements, finding heating-energy deviations of 8.5% and 6.9% and solar-thermal deviations of 0.4%, while meeting ASHRAE Guideline 14 limits.
- The paper provides consistently structured raw and kNN-imputed data with encoded metadata for benchmarking, MPC, reinforcement learning, fault diagnosis, transfer learning, and simulation-to-reality studies, while noting that proprietary models limit full reproducibility.
Motivation and positioning
Building operations account for roughly 30% of global energy consumption and 26% of energy-related greenhouse gas emissions, and methods such as model predictive control (MPC), reinforcement learning (RL), and automated fault detection and diagnosis (FDD) have demonstrated substantial potential for reducing this demand. All of these approaches depend on accurate models of building thermal dynamics and HVAC operation, yet most buildings lack the sensor infrastructure and historical data needed to train such models. Transfer learning—pretraining models on diverse source buildings before fine-tuning to a target building—is a practical remedy, but it requires datasets that span wide ranges of building characteristics, system configurations, occupant behaviors, and climates. Existing transfer learning studies in the building domain have largely been evaluated only on synthetic single-zone targets, and existing public datasets are poorly suited to thermal modeling: the Building Data Genome Project and BuildingsBench focus on electrical loads, ecobee provides thermostat data without detailed HVAC or multi-zone observations, the HOT dataset is purely synthetic at scale (~150,000 homes) without real-world operation, and IDEAL offers multi-zone sensing for 255 homes but with inconsistent measurements (only 32 of 255 homes include indoor temperatures), no weather data, and no heat pump systems.
The ThermBuild dataset addresses this gap by pairing measurements from two instrumented single-family homes with simulations of 958 TRNSYS building variants, all at 15-minute resolution. It is explicitly designed for multi-zone thermal dynamics modeling, transfer learning, generalization studies, simulation-to-reality (Sim2Real) transfer, benchmarking, and reproducible research.
Real-world measurement campaign
The measurement component derives from a 15-month campaign (7 February 2025 to 30 April 2026) conducted in the TwinHouses N2 and O5 at the Fraunhofer IBP site near Holzkirchen, Germany—full-scale test buildings previously used for empirical validation in IEA EBC Annexes 58 and 71. Each house provides 140 m² of living space (the German average for single-family homes), was constructed around 1980, has been retrofitted to meet GEG 2020 requirements for newly refurbished buildings, and contains seven common rooms with triple-pane insulating glazing. Basements were excluded and held at 18 °C.
The two houses differ deliberately in building service equipment (BSE). BSE1 (O5) uses an iDM iPump compact unit with only a 50 L buffer for defrost cycles, supplying the distribution system nearly directly; BSE2 (N2) comprises a wall-mounted indoor unit, an 825 L DHW storage tank supported by ~6 m² of solar thermal collectors, and a 500 L heating buffer. Together these represent direct-coupled versus storage-coupled heat pump architectures with distinct thermal dynamics. Both use identical air-source outdoor units and wet-screed (ground floor) / dry-screed (attic) underfloor heating. Occupancy is emulated via room-wise electrical heat gain simulators with approximately 50/50 convective-radiative split, driven by stochastic Occdem profiles; setpoints are 20 °C with no night setback, and mechanical ventilation runs at 155 m³/h.
Data acquisition used the Beckhoff PLC (1-second sampling) plus the iDM heat pump's internal logging, preferring calibrated PLC data where both existed. Residual gaps were filled with k-nearest-neighbor imputation (k=5); raw and imputed versions are published separately so users can evaluate imputation effects themselves—an important methodological transparency choice.
Simulation study design
The simulated corpus derives from validated TRNSYS 18 baseline models of both TwinHouses, varied systematically across ventilation type (2), location (5), size (3: 80%, 100%, 120%), building age (3: 1980, 2000, 2020), thermal mass (light/middle/heavy), glazing (standard vs. solar protection), and BSE configuration (2), yielding 1,080 unique configurations. Randomly assigned parameters include temperature setpoints (base 21 °C ± 2.5 K, with room-specific adjustments and night setback in 70% of cases), Occdem occupancy profiles (2–4 persons), orientation (45° increments), and room count (5–7). Airflows are modeled with TRNFLOW/COMIS. Of the 1,080 generated models, 122 failed convergence criteria (tolerance 0.1%, up to 400 iterations per 5-minute timestep, discarded if >0.3% of timesteps failed), leaving 958 successful runs totaling roughly 8.5 days of computation. Simulations ran at 5-minute resolution and were aggregated to 15-minute means for consistency with the measurements; each case spans three years of non-repeating Meteonorm 8 weather across Holzkirchen, Bolzano, Paris, Warsaw, and Copenhagen.
A notable design decision: the authors intentionally omitted parameter optimization/calibration of the baseline models beyond ASHRAE Guideline 14 compliance, because the dataset aims to represent a broader building stock rather than a tuned digital twin of the TwinHouses. This is a defensible but consequential trade-off—users seeking high-fidelity per-building models should be aware the simulated dynamics carry un-tuned simplifications.
Data records and structure
The dataset is distributed via Fraunhofer's FORDatis repository (DOI 10.24406/fordatis/445) as four archives: raw measurements, kNN-imputed measurements, supplementary vertical temperature stratification and mean radiant temperature data (not present in the simulated files), and the full simulation study. Each building is one CSV file with a uniform column layout covering:
| Domain |
Representative variables |
Heat pump (hp) |
Electrical power, compressor frequency, flow/return temperatures, thermal power, mass flow |
Distribution (dist) |
Buffer tank temperatures (top/mid/bottom), circuit mass flow, thermal output |
DHW (dhw) |
Tank temperatures, draw-off flow/temperature/thermal power |
Solar (solar) |
Collector loop flow temperatures and thermal gain |
Rooms (roomID) |
Air temperature, relative humidity, internal gains, setpoint, valve position, window state |
Weather (wea) |
Direct/diffuse horizontal irradiance, ambient temperature/humidity, wind speed/direction |
File names encode all metadata (age, mass, glazing, weather, size, ventilation, setback, occupancy profile, rotation, rooms), enabling direct filtering without separate metadata lookups. NaN conventions handle absent systems (e.g., BSE1 lacks DHW) and missing zones, and the paper correctly cautions that supply/return temperatures are physically meaningful only when mass flow exceeds zero—a detail that matters for downstream feature engineering.
Technical validation
Sensor quality assurance included co-location comparison of all room air temperature sensors under a common radiation shield, lab calibration of water temperature sensors, gravimetric validation of magneto-inductive flow meters repeated roughly every two months (with drift correction and eventual replacement by calibrated ultrasonic meters), certified calibration of ventilation airflow meters, and dual redundant acquisition systems.
Simulation validation followed Zhai et al.'s recommended practices. Against TwinHouse measurements over a 15-day March 2025 period spanning varied outdoor conditions, cumulative heating energy deviation was 8.5% for O5 and 6.9% for N2, within ASHRAE Guideline 14's 10% limit; solar thermal production deviated by only 0.4%. Plausibility checks then verified that every varied parameter produces its expected physical signature in three-year summary statistics, at both building and room level: older buildings show higher heating demand and lower COP ratios due to higher supply temperatures; Bolzano dominates cooling demand while Holzkirchen and Warsaw lead heating demand; solar protection glazing reduces cooling load without materially affecting heating; natural ventilation produces markedly lower minimum room temperatures and humidities than mechanical ventilation; and night setback lowers minimum room temperatures specifically in mechanically ventilated cases. These checks confirm correct implementation of the variation matrix rather than merely plausible-looking outputs.
Limitations and open questions
Several constraints should inform usage. The heat pump performance polynomials and control logic are proprietary and cannot be shared, and consequently the TRNSYS models themselves are not published—so the community receives outputs, not the generative models, limiting reproducibility of the simulation pipeline itself. The COP characteristics do not vary across building ages because units were scaled rather than re-specified, which flattens one axis of realistic heat pump diversity. The validation window against measurements covers only 15 days in March 2025, leaving seasonal generalization of the simulation baselines (particularly cooling-season behavior) empirically unverified. Airtightness is held constant across ventilation modes and building ages despite being strongly age-correlated in reality, and infiltration rates versus model assumptions are cited by the authors as a residual error source. The kitchen–living room merge used for six-room layouts is acknowledged as atypical. Finally, whether pretraining on this corpus yields measurable Sim2Real benefits when transferred to genuinely unseen occupied homes—as opposed to the well-instrumented TwinHouses—remains an open empirical question the dataset enables but does not answer.
Conclusion
ThermBuild contributes a carefully validated, consistently structured corpus of 960 multi-zone residential buildings—two measured, 958 simulated—with matched schemas, encoded metadata, and coverage of the principal drivers of residential thermal dynamics in Central Europe. Its paired real/synthetic design directly supports transfer learning, generalization, and Sim2Real research that prior electrical-load-centric or single-zone datasets could not accommodate, and its transparent handling of gaps, calibration, and convergence failures reflects rigorous data stewardship. The withheld simulation models and proprietary heat pump characterization are the principal restrictions on full reproducibility.