Enhancing building energy through regularized Bayesian neural networks for precise occupancy detection

Improving building energy efficiency is essential for promoting sustainable construction practices and minimizing operational costs. This study introduces a physics-based framework that integrates domain-specific constraints into a regularized Bayesian Neural Network (BNN) to enhance occupancy detection accuracy, an essential factor for optimizing building energy management. Unlike conventional approaches, our method leverages a unique combination of environmental sensor data (temperature, humidity, light, CO₂) and engineered features such as heating degree days (HDD) to improve predictive performance. Additionally, a physics-based regularizer is incorporated within the BNN model to ensure predictions adhere to the fundamental physical principles of the building, enhancing both reliability and uncertainty estimation. Tested on office building data from the University of Mons in Belgium, the proposed framework achieves 96–99 % accuracy across three test cases, outperforming traditional methods like Gradient Boosting Machine, Support Vector Machine, and Naïve Bayes. A user-friendly graphical interface was developed to facilitate real-world adoption, enabling facility managers, energy analysts, and building operators to seamlessly implement the approach without extensive technical expertise. By improving the precision of occupancy detection, this research supports more efficient HVAC control, enhanced occupant comfort, and substantial energy savings, an impact well-documented in previous studies that report potential reductions in energy consumption ranging from 20 to 30 %. The findings contribute to the advancement of intelligent building automation, offering a scalable solution for reducing carbon footprints and operational costs while promoting sustainable construction practices.