Implementing optimised HVAC setpoints via thermal adaptation behaviour recognition in residential buildings

The easiest way to improve energy efficiency and thermal comfort in existing buildings is to adjust setpoints. However, although Chinese standards regulated the heating and cooling setpoints, they lacked consideration for occupants’ thermal comfort. Therefore, this study aimed to identify optimal heating, cooling, and ventilation setpoints to enhance building energy performance and indoor thermal comfort in Chinese residential buildings. After obtaining these setpoints, it is necessary to adjust them based on real-time occupant thermal behaviour. Traditional comfort surveys fail to capture real-time insights, and conventional sensors lack the ability to record detailed occupancy data continuously. Deep learning and computer vision offer promising solutions, yet many existing models have low frame rates and high computational demands, potentially negating energy savings. To address this, this study developed a novel occupant thermal adaptation behaviour recognition model that balances accuracy, real-time performance and computational resource usage to enable effective operation indoors. This study used Design Builder to simulate the energy usage and thermal comfort with different setpoints. Typical cities representing China’s climate zones were selected via machine learning. Results indicated that adjusting setpoints in Harbin, Beijing, and Shanghai could increase comfortable days by 43.95%, 54.23%, and 23.36% respectively, with minimal impact on energy usage. In Guangzhou, energy usage could be reduced by 7.56% with a 5.91% decrease in comfortable days, while Kunming could see an 11.11% increase in comfortable days with a 5.92% rise in energy consumption. Additionally, proposed setpoint combinations demonstrated resilience to future weather conditions in most typical cities, except for Guangzhou. For behaviour recognition, a multi-camera Raspberry Pi 3B+ system captured a custom dataset of 400 videos from four angles, covering dressing, undressing, sitting, and standing. Compared to SlowFast (SF) and Spatial Temporal Graph Convolutional Networks (ST- GCN), the proposed lightweight skeletal temporal model achieved 0.975 accuracy on the Kungliga Tekniska Högskolan (KTH) dataset, running at 31.38 frames per second (FPS) on a Graphics Processing Unit (GPU)—over three times faster than ST-GCN and more than twelve times faster than SF—while using only 13.71% Central Processing Unit (CPU) and 33.05% GPU. When run on the CPU, it achieved 25.3 FPS with 56.10% CPU usage, proving its practicality for platforms without GPU support. When evaluated on the custom dataset, a double Long Short-Term Memory (LSTM) with an attention mechanism was introduced to better handle the increased action complexity, preserving a high accuracy of 0.963. Although the frame rate experienced a slight reduction compared to the results on the KTH dataset—dropping from 31.38 to 30.95 FPS on GPU and from 25.3 to 18.98 FPS on CPU—the model beneficially exhibited lower CPU and GPU usage, highlighting its potential for energy-efficient deployment in smart building applications. The model was further deployed on an NVIDIA Jetson Orin Nano, enabling stable long-term operation and supporting simultaneous multi-person recognition. Overall, this study selected the optimal HVAC setpoints for residential buildings in China and presents a practical, AI-driven solution for occupant thermal adaptation behaviour recognition, making it well-suited for energy-saving applications in buildings.