Machine learning-based adaptive control of PV shading for residential energy and visual comfort optimization

Building-integrated photovoltaic (BIPV) façade technology effectively reduces urban energy consumption. However, dynamic optimization strategies for photovoltaic shading devices (PVSDs), particularly in residential buildings, remain insufficiently studied. Using public housing in Hong Kong as a case study, this research investigates how building orientation and internal spatial functions influence shading performance. A random forest-based optimization strategy was developed to simultaneously enhance energy efficiency and visual comfort, with Shapley values quantifying key influencing factors. Results showed that the random forest algorithm achieved over 85 % prediction accuracy (±5° tolerance), outperforming eight other algorithms. Orientation-specific strategies emerged: south façades responded primarily to solar azimuth angles with moderate adjustments (15°–50°), whereas west façades required significant adjustments (20°–80°) due to rapid solar altitude changes. Three distinct control strategies were compared: a typical-day-based approach achieved energy savings of 37.01 % (south) and 29.78 % (west), while a machine learning predictive control strategy best balanced energy savings (>30 %) and glare reduction (annual glare-free hours: 90.95 % south; 88.36 % west). These findings provide architects and policymakers with actionable insights into implementing adaptive PVSD controls, thus facilitating energy-efficient and comfortable residential environments in subtropical urban areas.