A Machine Learning-Based Thermal Comfort Prediction Model for Older Adults

Abstract

Indoor thermal comfort plays a critical role in safeguarding the health and quality of life of older adults (aged 60 and above), whose thermoregulatory capacity is diminished compared with younger populations due to reduced basal metabolic rate, impaired vasomotor responses, and elevated thermal sensation thresholds. While the Predicted Mean Vote (PMV) model has long served as a benchmark for thermal comfort assessment, its parameters are calibrated predominantly on young, healthy subjects and thus produce systematic prediction biases for older adults. Data-driven machine learning approaches have emerged as promising alternatives; however, existing studies that advocate deep learning architectures for thermal comfort prediction are predominantly based on small-sample datasets (N<500), leaving their generalizability under large-scale data conditions unverified. This study systematically compares a Convolutional-Recurrent Neural Network (CRNN, comprising one-dimensional convolution and Gated Recurrent Unit layers) with three established machine learning algorithms—K-Nearest Neighbors (KNN), Gradient Boosting Decision Trees (GBDT), and Random Forest (RF)—using a dedicated dataset of 5 820 older adult samples constructed from the ASHRAE Global Thermal Comfort Database II. Model performance was evaluated across both regression (continuous thermal sensation prediction) and classification (discrete seven-level thermal sensation determination) tasks. Results demonstrate that RF outperformed all models: in regression, RF achieved a mean absolute error (MAE) of 0.597 3 and root mean square error (RMSE) of 0.842 7, approximately 13.1% and 15.1% lower than CRNN (MAE = 0.687 6, RMSE = 0.992 2), respectively. In classification, RF attained 78.0% accuracy and a weighted F1 score of 0.539, compared with 68.7% and 0.513 for CRNN. The CRNN exhibited pronounced overfitting and majority-class bias, attributable to an architectural incompatibility between its sequential modeling design and the non-temporal tabular nature of thermal comfort data. These findings provide evidence-based guidance for selecting models to construct individualized thermal comfort prediction systems for older adults and inform the development of intelligent age-friendly textile products.