Improved Thermal Comfort Model Leveraging Conditional Tabular GAN Focusing on Feature Selection

Occupants’ personal thermal comfort (PTC) is indispensable for their well-being, physical and mental health, and work efficiency. The heating system controlled by Artificial intelligence (AI) can calibrate the indoor thermal condition automatically by analyzing different physiological and environmental variables. Predicting occupants’ personal thermal comfort preferences in a smart home can be a prerequisite to adjusting the indoor temperature that might provide a comfortable environment. Modeling personal thermal comfort preference is challenging due to two major challenges including the inadequancy of the data and it’s high dimensionality. Adequate amount of data is an obvious requirement for training efficient machine learning (ML) models. In addition, high-dimensional data tends to have multiple features that are irrelevant, noisy and might hinder ML models’ high performance. To this end, we proposed a personal thermal comfort preference prediction framework by employing synthetically generated data introducing generative adversarial network (CTGAN) and multiple feature selection techniques. We first address the data inadequacy challenge by applying CTGAN to generate synthetic data which considers challenges associated with multimodal distributions and categorical features. Then multiple feature selection techniques are incorporated to identify the best possible sets of features from high-dimensional feature sets. The wide range of experimental settings on a standard dataset demonstrated the state-of-the-art performance in predicting personal thermal comfort preference. The results clearly indicate that the ML models trained on synthetic data achieved significantly better performance than the models trained on real data. In turn, our methods with supervised feature selection techniques achieved higher performance in terms of evaluation metrics including accuracy, Cohen’s kappa, and area under the curve (AUC) outperforming conventional methods. Additionally, our method enhances the explainability of the thermal preference prediction system and provides an avenue for thermal comfort experiment designers to consciously select data to be collected.