A hybrid learning approach to model the diversity of window-opening behavior

The diverse window-opening behaviors of individuals can result in significant differences in indoor thermal environments, air quality, and energy utilization. However, the majority of existing studies focus on constructing an average window operation model, thus overlooking the diversity of behaviors. Current methods for addressing behavioral diversity face challenges with integration into building performance simulation software and are highly dependent on data scale. To address these limitations, this study proposes a novel approach that combines unsupervised learning (K-Means) and supervised learning (Light Gradient Boosting Machine, LightGBM) for modeling the diverse window-opening behaviors. Furthermore, the SHapley Additive exPlanations (SHAP) was employed to interpret the predictive model. This study yielded four key findings: 1) There were 12 different window-opening behavior patterns. Interestingly, 65% of the residents’ window-opening behaviors were not influenced by environmental factors but were instead a matter of personal habit. 2) Using random sampling to divide the dataset may pose a risk of data leakage. The time series cross-validation method is more suitable for evaluating the performance of the window state prediction model. 3) Under the time series sampling strategy, the LightGBM model incorporating behavioral diversity improved the prediction accuracy by 1.3%-10.4% compared to the standalone LightGBM model. Notably, when the daily average window opening time was used as a clustering feature in the LightGBM model (Cluster(T)-LightGBM), the accuracy reached 87.1%. 4) The SHAP feature analysis highlighted high-intensity window-opening categories, outdoor temperature, and indoor CO2 concentration as the most pivotal predictors.