Learning personalized visual comfort: A data-driven adaptive shading control framework using artificial intelligence

Personalized shading control is critical for optimizing occupant visual comfort and building energy efficiency, yet conventional systems rely on static and generalized settings. This study proposes a machine learning (ML) framework that infers individual visual discomfort thresholds from occupants’ real-time interactions with smart blinds. A comparative evaluation of multiple models, including random forest (RF) and long short-term memory (LSTM), found that the convolutional neural networks (CNN) were most effective. Trained on a comprehensive dataset simulating 4860 unique occupant profiles, the CNN leverages a spatiotemporal reconstruction of multi-day behavioral data. In simulations, the CNN-based control system reduced behavioral mismatch scores by approximately 70 % and increased the occupant acceptability rate from 68 % to 83 % compared to a conventional default strategy. The framework’s real-world applicability was confirmed in a preliminary on-site test, where the model successfully learned a user’s preferences from a short observation period and significantly improved system performance. The findings show that model accuracy is highest for metrics that most frequently trigger overrides, such as Daylight Glare Probability (DGP), demonstrating the system’s ability to identify and respond to the most critical comfort drivers. This study presents a robust simulation-based methodology for developing occupant-centric shading systems, which shows promising results in preliminary on-site testing and has the potential to enhance indoor environmental quality and support more energy-efficient building operations.