Tien, Paige Wenbin
(2023)
Deep learning-powered vision-based energy management system for next-gen built environment.
PhD thesis, University of Nottingham.
Abstract
Heating, ventilation and air-conditioning (HVAC) systems provide thermally comfortable spaces for occupants, and their consumption is strongly related to how occupants utilise the building. The over- or under-utilisation of spaces and the increased adoption of flexible working hours lead to unnecessary energy usage in buildings with HVAC systems operated using static or fixed schedules during unoccupied periods. Demand-driven methods can enable HVAC systems to adapt and make timely responses to dynamic changes in occupancy. Approaches central to the implementation of a demand-driven approach are accurate in providing real-time information on occupancy, including the count, localisation and activity levels. While conventional occupancy sensors exist and can provide information on the number and location of occupants, their ability to detect and recognise occupancy activities is limited. This includes the operation of windows and appliances, which can impact the building’s performance.
Artificial intelligence (AI) has recently become a critical tool in enhancing the energy performance of buildings and occupant satisfaction and health. Recent studies have shown the capabilities of AI methods, such as computer vision and deep learning in detecting and recognising human activities. The recent emergence of deep learning algorithms has propelled computer vision applications and performance. While several studies used deep learning and computer vision to recognise human motion or activity, there is limited work on integrating these methods with building energy systems. Such methods can be used to obtain accurate and real-time information about the occupants for assisting in the operation of HVAC systems.
In this research, a demand-driven deep learning framework was proposed to detect and recognise occupancy behaviour for optimising the operation of building HVAC systems. The computer vision-based deep learning algorithm, convolutional neural network (CNN), was selected to develop the vision-based detector to recognise common occupancy activities such as sitting, standing, walking and opening and closing windows. A dataset consisting of images of occupants in buildings performing different activities was formed to perform the training the model. The trained model was deployed to an AI-powered camera to perform real-time detection within selected case study building spaces, which include university tutorial rooms and offices.
Two main types of detectors were developed to show the capabilities of the proposed approach; this includes the occupancy activity detector and the window opening detector. Both detectors were based on the Faster R-CNN with Inception V2 model, which was trained and tested using the same approach. In addition, the influence of different parameters on the performance, such as the training data size, labelling method, and how real-time detection was conducted in different indoor spaces was evaluated. The results have shown that a single response 'people detector’ can accurately understand the number of people within a detected space. The ‘occupancy activity detector’ could provide data towards the prediction of the internal heat emissions of buildings. Furthermore, window detectors were formed to recognise the times when windows are opened, providing insights into the potential ventilation heat losses through this type of ventilation strategy employed in buildings. The information generated by the detector is then outputted as profiles, which are called Deep Learning Influence Profiles (DLIP).
Building energy simulation (BES) was used to assess the potential impact of the use of detection and recognition methods on building performance, such as ventilation heat loss and energy demands. The generated DLIPs were inputted into the BES tool. Comparisons with static or scheduled occupancy profiles, currently used in conventional HVAC systems and building energy modelling were made. The results showed that the over- or under-estimation of the occupancy heat gains could lead to inaccurate heating and cooling energy predictions. The deep learning detection method showed that the occupancy heat gains could be represented more accurately compared to static office occupancy profiles. A difference of up to 55% was observed between occupancy DLIP and static heat gain profile. Similarly, the window detection method enabled accurate recognition of the opening and closing of windows and the prediction of ventilation heat losses.
BES was conducted for various scenario-based cases that represented typical and/or extreme situations that would occur within selected case study buildings. Results showed that the detection methods could be useful for modulating heating and cooling systems to minimise building energy losses while providing adequate indoor air quality and thermal conditions. Based on the developed individual detectors, combined detectors were formed and also assessed during experimental tests and analysis using BES.
The vision-based technique’s integration with the building control system was discussed. A heat gain prediction and optimisation strategy were proposed along with a hybrid controller that optimises energy use and thermal comfort. This should be further developed in future works and assessed in real building installations. This work also discussed the limitations and practical challenges of implementing the proposed technology. Initial results of survey-based questionnaires highlighted the importance of informing occupants about the framework approach and how DLIPs were formed. In all, preference is towards a less intrusive and effective approach that could meet the needs of optimising building energy loads for the next-gen built environment.
Actions (Archive Staff Only)
 |
Edit View |
Tools
Tools
