A Personalised Thermal Comfort Model Using a Bayesian Network

Frederik Auffenberg, Sebastian Stein, Alex RogersIJCAI 2015

Introduction

Smart thermostats often fail to provide optimal comfort because they rely on generalized models that do not account for individual variability. This paper proposes a personalized thermal comfort model based on a Bayesian Network that learns user preferences from minimal feedback, significantly outperforming industry standards like ASHRAE 55.

Problem Statement

Standard thermal comfort models, such as Fanger’s Predicted Mean Vote (PMV) and adaptive models, are designed for groups in shared spaces. They often ignore individual sensitivities to humidity, seasonal psychological shifts, and personal "neutral" temperatures. This leads to energy waste and user discomfort in residential settings where individual preferences vary wildly.

Methodology

The authors developed a hybrid model represented as a Bayesian Network (specifically a factor graph). This model simplifies complex heat-balance physics to use accessible sensor data: indoor operative temperature, relative humidity, outdoor temperature, and the date.

The Bayesian Network model architecture, showing latent variables and factors used for probabilistic inference.

Key features include:

  • Latent Variable Inference: Individual traits like base comfort temperature ($T^*$) and thermal sensitivity ($\gamma_v$) are treated as latent variables to be inferred from behavior.
  • Probabilistic Learning: Using the Infer.NET library and Expectation Propagation (EP), the model updates its beliefs with each piece of user feedback (e.g., adjusting the thermostat).
  • Dual Output: The system predicts both the current thermal sensation "vote" and the user's "optimal temperature," allowing for energy-efficient temperature "drift" within a personalized comfort range.

Evaluation and Results

The model was tested using the ASHRAE RP-884 dataset, covering 553 individuals across diverse environments, including Pakistan (naturally ventilated), Athens (HVAC offices), and San Francisco. Performance was measured using the Root Mean Square Error (RMSE) of the predicted thermal sensation votes.

RMSE comparison across different datasets. The personalized model (red) consistently outperforms baselines and converges quickly, typically within 10 observations.

The personalized model consistently outperformed both PMV and Adaptive models. On average, it achieved a 23.5% accuracy gain over PMV and 17.5% over Adaptive models. A critical finding was the convergence rate: the model achieves high accuracy after approximately 10 observations of user feedback. In Athens, the model reduced prediction error by 30% compared to standard baselines.

Conclusion

By prioritizing individual personalization over complex group-based physics, the proposed Bayesian Network model provides a more accurate and responsive solution for smart climate control. While challenges remain in multi-occupancy spaces, this approach demonstrates that thermostats can become significantly "smarter" by learning directly from the people they serve.

Summary updated

BibTeX
@inproceedings{auffenberg2015personalised,
  author = {Frederik Auffenberg and Sebastian Stein and Alex Rogers},
  title = {{A Personalised Thermal Comfort Model Using a Bayesian Network}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2015},
}
Figure