Graph Neural Network Ensemble with Particle Swarm Optimization for Privacy-Preserving Thermal Comfort Prediction

Abstract

Heating, ventilation, and air conditioning (HVAC) systems account for nearly 60% of energy consumption in commercial buildings, yet occupant dissatisfaction with thermal comfort remains high. To address this challenge, we propose a novel framework that leverages the ASHRAE Global Thermal Comfort Database II to predict individual thermal preferences while ensuring energy efficiency. Unlike prior deep learning approaches, our method employs a Graph Neural Network (GNN) ensemble with attention mechanisms, enabling the model to capture complex relationships among personal, environmental, and contextual variables across seasons and building types. Feature selection is performed using Particle Swarm Optimization (PSO), which enhances diversity and avoids premature convergence by dynamically updating particle velocities and positions. The selected features are then fed into the GNN ensemble, which integrates multiple graph-based learners to improve robustness. Hyperparameter tuning is conducted using Bayesian Optimization, balancing exploration and exploitation to identify optimal learning rates, dropout ratios, and batch sizes. Experimental results on the ASHRAE dataset demonstrate that the proposed GNN-PSO-Bayesian framework achieves 96.8% accuracy, outperforming traditional classifiers while maintaining interpretability and scalability. This architecture highlights the potential of graph-based learning for occupant-centric thermal comfort prediction, offering a pathway toward sustainable and adaptive HVAC control.