COMBINING MACHINE LEARNING AND BAYESIAN MODELS FOR PREDICTING SHADING ADJUSTMENTS IN OFFICE BUILDINGS

ABSTRACT With the growing demand for smart buildings, optimizing shading adjustments in office environments is crucial for enhancing energy efficiency and occupant comfort. This study presents a dual-layer model combining machine learning and Bayesian methods to predict shading adjustment behavior in office buildings, addressing the complexity often overlooked by previous approaches. Using long-term monitoring data of indoor and outdoor environmental parameters, occupant presence, and shading statuses, the model categorizes windows into three types—far-range, midrange, and close-range—based on the complexity of adjustments, and trains models for each category. Sensitivity analysis identifies key factors influencing model performance. Multiple algorithms, including Decision Tree, Random Forest, Gradient Boosting, Multilayer Perceptron, and K-Nearest Neighbors, were evaluated through ROC curve analysis, with Random Forest emerging as the most suitable. The model’s performance, validated through cross-validation, confusion matrices, and various metrics, demonstrates high accuracy in predicting shading behavior. The findings underscore the potential of integrating machine learning and Bayesian models to improve shading control in smart buildings, offering a robust tool for enhancing energy management and occupant comfort.