Analysis of Post-Retrofit Changes in Window Operation Behavior in Retrofitted University Buildings Using Machine Learning

Building energy retrofits are a key strategy for reducing energy consumption in the existing building stock; however, the success of this initiative is dependent on occupant behavior after the retrofit. Specifically, changes in window openings may alter ventilation patterns and influence the anticipated energy savings. This study focuses on analyzing the change in window operation behavior in post-retrofit buildings using three machine learning models. The data analyzed were obtained from an open-access source and collected from eight residential dormitories on a university campus in Syracuse, New York, USA. Pre-retrofitting data were captured with a 1-minute time resolution between August 2021 and May 2022. Post-retrofitting data were obtained at the same resolution between August 2022 and May 2023. Supervised machine learning models, such as Random Forest, XGBoost, and Gradient Boosting, were used to predict window-opening events and investigate the differences in the behavior of occupants in the pre- and post-retrofit phases using feature-importance analysis. Window-opening frequency varied significantly across the eight buildings rather than showing a consistent trend. While five buildings experienced reductions of 21–81%, the other three saw increases of 21– 78%, resulting in an average reduction of 17.5%. This shift is accompanied by a change in the temperature thresholds that initiate the opening of the window, which also decreases by a mean of 1.8 °C. The patterns of feature importance indicate a shift in behavior from being dominated by indoor temperature before retrofitting to being dominated by indoor air quality and the suitability of outdoor conditions for ventilation. The Gradient Boosting models demonstrated moderate and building-specific predictive performance, with area under the curve (AUC) values between 0.79 and 0.82 for window state prediction across buildings, reflecting heterogeneous behavioral patterns. The findings indicate that behavioral adaptation effects should be considered to enhance the performance of retrofit assessment and building energy modelling to enhance the quality of projected energy savings. Feature Importance Analysis o