Work

Spatiotemporal deep learning for multimodal thermal comfort assessment and prediction in community elderly day-care centers

Year: 2026

Type: article

Abstract: Thermal comfort management in community elderly day-care centers requires models that reflect the physiology, routines, and spatial use patterns of older occupants rather than extrapolating from standard office-based assumptions. This study develops a SpatioTemporal Convolutional Neural Network–Long Short-Term Memory–Model-Agnostic Meta-Learning (CNN–LSTM–MAML; ST-CLM) framework for zone-level thermal comfort assessment using environmental measurements, non-contact infrared skin temperatures, spatial coordinates, and modified seven-point thermal sensation votes. A twelve-month field campaign in three day-care centers in Zhengzhou, China produced 13,500 valid observations from 247 participants aged 60 years and above. On the held-out test set, ST-CLM achieved 91.3% classification accuracy and a macro-averaged F1 score (the harmonic mean of precision and recall) of 0.896, exceeding support vector machine (75.4%), random forest (79.2%), and non-meta CNN–LSTM (86.7%) baselines. Under an exploratory five-window support setting (50 time-step feature rows per task rather than five isolated observations), the model-agnostic meta-learning component yielded a protocol-specific point estimate of 84.2% accuracy; this value is not presented as a general guarantee of five-observation calibration. Shapley additive explanations (SHAP) analysis ranked air temperature, forehead skin temperature, and carbon dioxide (CO 2 ) concentration as the leading predictors; using this three-feature subset preserved 95.4% of full-model accuracy. Model-generated spatial comfort maps showed pronounced intra-building variation, including entrance-lobby thermal sensation votes 1.6 units lower than activity-room values during the heating season. Relative to the predicted mean vote (PMV), which overestimated comfort by 1.23 units among participants aged 80 and above, ST-CLM reduced the error to 0.27 units. The results indicate that interpretable spatiotemporal meta-learning can support personalized and zone-differentiated heating, ventilation, and air-conditioning (HVAC) operation in age-friendly community care facilities.

Source: Energy and Buildings

DOI: 10.1016/j.enbuild.2026.118266

Authors: Quan Wen, Mazran Ismail, Muhammad Hafeez Abdul Nasir, Yanting Wu

Institutions: Universiti Sains Malaysia, Henan Industry and Trade Vocational College

Language: English

Topic: Building Energy and Comfort Optimization

Subfield: Building and Construction

Field: Engineering

Domain: Physical Sciences

SDG: Good health and well-being

Open Access status: hybrid

Estimated APC paid: $4,260