Modeling occupant air-conditioning behavior in Mediterranean offices with manual HVAC control: a case study conducted in Montpellier during summer 2025

Occupant behavior related to air conditioning (AC) has a considerable impact on the energy consumption of office buildings, particularly in southern France, which is characterized by a hot Mediterranean climate and frequent heat waves. Although several studies have developed behavioral models using logistic regression, relatively few have examined the application of machine learning methods to predict occupant behavior in French offices during the summer months. To address this gap, this paper presents a comparative study of four modeling approaches for predicting AC activation behavior: logistic regression, random forest, Extreme Gradient Boosting (XGBoost), and support vector machines. The research is an exploratory, case-study-based investigation carried out in Montpellier during the summer of 2025 (June to September) in six offices occupied by 14 individuals. Models were trained per office using datasets including 1270 to 3039 AC on/off observations, with class distributions ranging from 24–65% for the “on” status and 35–76% for the “off” status.
Five distinct occupant profiles were identified based on AC usage patterns. The highest F1-score achieved was 0.65, obtained with both logistic regression and random forest in an office with a single, consistent usage pattern. Lower performance was observed in offices with diverse behaviors and class imbalance. Compared with prior literature reporting F1-scores above 0.8, these results highlight the challenges of modeling AC behavior in small-scale, diverse office environments.
Overall, the findings show that model transferability is limited across different offices and that feature importance varies across algorithms, emphasizing the need for tailored, context-specific modeling approaches to predict AC activation behavior.