Abstract
[Display omitted]
•Adaptive Random Forest overcomes erratic occupancy forecasting.•Online learning provides robustness against holidays and sudden events.•Balances exceptional accuracy (>0.9) with outstanding computational speed.•A practical blueprint for next-generation smart building management.
Enhancing energy efficiency is critical for reducing the carbon footprint of the built environment, with accurate building occupancy prediction serving as a foundation for smart energy management. Traditional occupancy forecasting models often suffer from poor adaptability to sudden changes, such as holidays, and are limited by computational intensity, hindering their real-world scalability. To address these limitations, we propose a novel approach integrating a robust machine learning model with a dynamic re-training framework. We first benchmarked several models using high-frequency (1-minute interval) occupancy data from a meeting room, including simpler methods (e.g., Random Forest, Support Vector Regression) and advanced time-series techniques (e.g., Long Short-Term Memory, Temporal Convolutional Networks). The Random Forest (RF) model was selected for its superior balance of prediction performance and computational efficiency. To ensure robustness and adaptability, the RF model was augmented with time-based, error-based, and event-based online learning mechanisms. This integration allows the model to dynamically adapt to occupancy fluctuations and abrupt changes, such as holiday periods. This combined online RF model achieved a prediction accuracy R2 of 0.92, representing an 80% improvement over the base RF model’s performance in handling abrupt events. These findings demonstrate the significant potential of integrating advanced machine learning with dynamic re-training to create highly adaptive, accurate, and energy-efficient building management systems.