Personalizing Individual Comfort in the Group Setting

Emil Laftchiev, Diego Romeres, Daniel NikovskiAAAI 2021 · DOI

Introduction

Maintaining thermal comfort in shared office spaces is a persistent challenge for building management. Traditional HVAC systems typically operate based on a single temperature sensor and a fixed setpoint, ignoring the subjective nature of human comfort. Factors such as metabolism, clothing, and personal preference mean that a temperature ideal for one person may be uncomfortable for another, often leading to "thermostat wars" and energy inefficiency.

This paper presents a data-driven framework designed to personalize thermal comfort in group settings. By combining weakly-supervised neural networks to learn individual preferences with an autoencoding approach to model room thermodynamics, the system can regulate local temperatures to within 0.5°C of a user’s ideal preference, significantly outperforming traditional fixed-setpoint methods.

Methodology

The proposed architecture addresses the complexity of group comfort through a three-part system designed for ease of deployment and minimal user burden.

Weakly-Supervised User Modeling

To avoid the "cold start" problem and the need for extensive manual ratings, the system initializes with a "standard" comfort model based on engineering standards (such as ASHRAE). This model is implemented as a neural network where hidden neurons represent the boundaries of comfort (e.g., too hot, too cold). Personalization occurs only when a user interacts with the thermostat; these interactions are treated as labels of discomfort, triggering backpropagation to refine the individual's specific comfort zone.

Comparison of standard and personalized comfort models on a psychrometric chart, showing how the system narrows the broad standard comfort zone to a specific individual preference.

Autoencoding Room Thermodynamics

To understand how HVAC setpoints affect specific locations in a room, the authors use an autoencoder. The "encoder" learns an inverse model (mapping room states to optimal HVAC setpoints), while the "decoder" learns a forward model (predicting room states from setpoints). This "black-box" approach captures non-linear micro-climates—such as the difference between a desk near a window and one under a vent—without requiring expensive manual calibration of physical equations.

Control Law Learning

The system determines final HVAC settings using two primary methods:

  • Retuned Inverse Model (RIM): Warm-starts with the learned room model and is fine-tuned to minimize the gap between predicted local temperatures and user preferences.
  • Tuned Linear Map (TLM): Adds a simple linear layer to allow for rapid adaptation with very few data samples.

Experimental Validation

The framework was validated through a 10-day field experiment in an open office in Japan during August 2019. The setup included 5 HVAC units and 39 sensors providing 78 measurements at 1-minute intervals. Over the study period, 18 participants generated 12,425 data points and 136 manual setpoint change requests.

Key Results and Insights

The system demonstrated high precision and computational efficiency:

  • High Precision: On average, the system regulated the temperature at each user's specific location to within 0.5°C of their desired temperature.
  • Superior Performance: The RIM and TLM methods outperformed standard linear regression and fixed-average setpoint baselines by a factor of at least two in terms of Root Mean Squared Error (RMSE).
  • Edge-Friendly Computation: Once trained, the control laws can determine optimal settings in approximately 3.7ms, making the approach suitable for low-power IoT controllers.
Histograms showing the distribution of RMSE and MAE for the Retuned Inverse Model (RIM) compared to individual optimization, highlighting the system's high precision.

A key insight is the importance of the "standard" starting point, which ensures occupants remain comfortable while the AI learns. Additionally, the inclusion of HVAC state data (like fan speed) in the neural network significantly improved prediction accuracy compared to models using only temperature and humidity data.

Limitations and Future Work

While the results are promising, the study was limited to a 10-day summer period. Future work is required to evaluate performance during seasonal transitions and to incorporate dynamic, time-dependent modeling to handle rapid changes in occupancy. Furthermore, the authors suggest exploring the social dynamics of group comfort, such as weighting preferences based on organizational roles or collective consensus.

Summary updated

BibTeX
@inproceedings{laftchiev2021personalizing,
  author = {Emil Laftchiev and Diego Romeres and Daniel Nikovski},
  title = {{Personalizing Individual Comfort in the Group Setting}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2021},
  doi = {10.1609/aaai.v35i17.17801},
}
Figure