Motivation and research question

Buildings account for around 40 % of final energy consumption and 36 % of CO₂ emissions in the EU (European Commission, 2020). At the same time, people spend up to 90% of their lives indoors (Brasche et al., 2005; Klepeis et al., 2001), where environmental quality affects health, well-being and productivity (Melikov, 2015; Awada et al., 2021; Crook & Burton, 2010; Wargocki et al., 2019). Building operations therefore face a dual challenge: reducing energy consumption and emissions without compromising occupant comfort. Established comfort standards such as PMV are based on laboratory studies and target the average thermal perception of a group, but fail to capture individual differences adequately. The PersonAI project investigated whether data-driven, personalised comfort models can resolve this tension and what effects a cohort-based temperature control strategy has on comfort and energy consumption under real building operating conditions.

Initial situation / status quo

Personal Comfort Models (PCM) use subjective feedback, physiological signals and environmental data to predict individual thermal preferences. However, previous PCM research has been largely limited to tropical climates and controlled laboratory settings with small sample sizes. Cohort Comfort Models (CCM) extend this approach by grouping occupants with similar comfort profiles, enabling personalisation even for new users with little or no historical data. Whether these approaches work under real-world conditions in a European climate zone, and what effect they actually achieve in building operations, had not been investigated prior to this project.

Project contents and objectives

PersonAI pursued four main objectives: defining requirements and use cases for user-centered energy services together with stakeholders, conducting an empirical longitudinal study to collect comfort data under everyday conditions, developing and testing Cohort Comfort Models with a subsequent proof of concept in a real building, and implementing open, privacy-compliant research data management.

Methodical procedure

The empirical study was conducted with 83 participants over four weeks in spring 2025. Data collection combined Ecological Momentary Assessment (seven Right-Here-Right-Now surveys per day via the Cozie app on the Apple Watch), physiological measurements (heart rate via Apple Watch and Oura Ring) and environmental data (temperature, humidity, CO₂ via stationary AirCO2ntrol sensors).

In addition, demographic, psychological and personality-related characteristics were collected through an onboarding survey. After data preparation, the final dataset comprised 72 participants. For comfort modelling, the published Cohort Comfort Models codebase by Quintana et al. (2023) was adopted and adapted to the project's own dataset. Temperature setpoints and comfort bands were derived descriptively from the resulting cohorts and deployed as setpoints in the proof of concept at the Living Labs of project partner Forschung Burgenland. The energy implications were evaluated through simulations using a data-driven model predictive control approach (DMPC) on two buildings with different construction types and technical equipment.

Results and conclusions

The comfort responses were highly imbalanced. The class "No Change" dominated for most participants. The trained comfort models were unable to reliably identify the less frequent classes "Cooler" and "Warmer". In the proof of concept, the cohort-based temperature control showed no clear comfort effect. A comparable increase in satisfaction was observed even for participants deliberately assigned to the wrong cohort.

A replication study conducted after the PoC on both the original data (Dorn dataset) and the PersonAI dataset identified methodological issues in the reference codebase. The evaluation strategy was replaced by a person-level Leave-One-Out Cross-Validation, which provides less biased estimates for small datasets and an information leak in the feature selection was corrected. The original evaluation metric F1-Micro was supplemented with F1-Macro, which weights classification performance equally across all three comfort classes. F1-Macro reveals that the seemingly good performance under F1-Micro is largely attributable to correct prediction of the dominant class "No Change". Even on the more extensive reference dataset, no statistically robust performance difference between CCMs and a simple baseline could be established. A comparison of correct and deliberately incorrect cohort assignments further showed that the reported performance gains are driven by a small number of outliers with near-homogeneous response patterns. The assumption that 60 data points per person are sufficient for stable comfort predictions could likewise not be confirmed.

The energy simulations conducted in parallel show, however, that wider comfort corridors in combination with model predictive control offer considerable savings potential. At the ENERGETIKUM, a building with controllable shading and thermally activated building systems, DMPC reduced heating energy demand by up to 47 % and the electrical energy demand of the heat pump by up to 45 % compared to a conventional two-point controller. At the LOWERGETIKUM, a building without controllable shading, savings were more modest at 7 to 8 %. The energy savings result primarily from the predictive control strategy and the thermal degrees of freedom it exploits, not from the personalisation of the comfort models.

Outlook

The project results are being pursued in several follow-up projects. The WellFit project replicates and extends the study design under summer conditions. Further planned projects aim to deepen the connection between indoor air quality, wearable data and AI-based analysis. The replication study will be published as an independent scientific paper. On the energy side, it is recommended that comfort bands should in future be treated not only as fixed constraints but as independent optimisation parameters, and that their effects should be evaluated on a building-specific basis, taking into account construction type, thermal mass, shading and available building services.

Project management

Technische Universität Graz

Project or cooperation partners

• DILT Analytics GmbH
• Forschung Burgenland GmbH
• Universität Graz - Institut für Öffentliches Recht und Politikwissenschaft

Gerald Schweiger
Inffeldgasse 16b/II
A-8010 Graz
Tel: +43 (316) 873 - 5747
E-Mail: gerald.schweiger@tugraz.at
Web: https://www.tugraz.at/institute/ist/home/