iBTEC: A tiny deep learning approach for indoor temperature and occupancy prediction in Smart Buildings

Maintaining thermal comfort while ensuring energy efficiency is a critical challenge in modern buildings. Smart Buildings (SBs) address this issue by integrating interconnected sensors and Internet of Things (IoT) technologies to automate systems such as heating, ventilation, air conditioning (HVAC), lighting, and water heating. However, achieving optimal HVAC control demands precise predictive modeling to effectively balance occupant comfort and energy consumption. In response, this paper introduces the Intelligent Building Thermal-Energy Controller (iBTEC), an innovative solution utilizing a Temporal Convolutional Network (TCN) enhanced by modern pruning techniques. The proposed iBTEC simultaneously predicts indoor temperature and occupancy, treating temperature prediction as a regression problem and occupancy prediction as a classification problem. To facilitate efficient deployment at the network edge and align with the principles of Green AI, this work employs two pruning strategies aimed at producing a lightweight model suitable for resource-constrained environments. These pruning techniques are justified as a means to make the approach greener, reducing computational overhead while maintaining accuracy. A prototype of the iBTEC was developed and tested in the offices of A.E.I Srl in Italy, leveraging real-time multivariate sensor data. Extensive evaluations were conducted using multiple performance metrics. Furthermore, the iBTEC approach was tested against several baseline TCN architectures, emphasizing reductions in model complexity, inference time, and training duration to identify suitable candidates for edge computing in SBs. Additionally, the robustness and reliability of the iBTEC approach were validated by using cross-validation and by using a state-of-the-art dataset, confirming its suitability for practical applications. Experimental results demonstrate that the proposed iBTEC model improves prediction accuracy compared with baseline deep learning models while significantly reducing computational complexity through pruning. In particular, hybrid pruning achieves up to 51.6% reduction in parameters, 52% reduction in arithmetic operations, and inference time 33% improvement compared to the baseline architecture, while maintaining comparable or improved prediction performance across both indoor temperature and occupancy tasks.