Reported as the largest energy consumer in recent years, buildings use a notable amount of the total energy produced in a given region: about 40% in Europe, 28% in China, and 39% in the UK. To save energy, many researchers have suggested focusing on heating, ventilation, and air-conditioning (HVAC) systems, since they consume the largest portion of energy across all types of buildings, especially commercial structures. In Hong Kong, commercial buildings consume about 42% of the total energy produced, and HVAC systems are the largest user. In a study conducted by Pérez-Lombard et al., the researchers found that HVAC systems accounted for a large percentage of energy usage in commercial buildings: 48% in the US, 55% in the UK, and 52% in Spain. The US Department of Energy recognizes HVAC systems as the most critical component in building energy efficiency assessment.

Many studies suggested that actual energy performances of buildings severely deviate from their original design conditions due to incorrect estimation of occupancy behavior. Significant discrepancies have been observed due to the complicated interrelationship of energy consumption in building facilities and occupancy behavior. Masoso and Grobler studied the electricity consumption in an office building and found 56% was consumed during non-working hours due to leaving on lightings or other devices when offices were unoccupied. Dong and Andrews implemented sensor-based occupancy behavior model and prediction method in building energy and comfort management systems and simulations suggested potential energy saving of 30% when compared with other basic energy savings by HVAC control strategies. Wood and Newborough found an energy reduction of 10% to 20% when studying the effect on real energy consumption by occupancy feedback method. The International Energy Agency (IEA) Energy in Building and Community (EBC) Programme Annex 66 has highlighted and concluded the significant roles of occupancy behavior in building performance study. It emphasizes that occupancy behavior is a key factor in the evaluation of energy-saving technology by observing how occupants understand and interact with the technology when the building is in use. Therefore, researchers have developed various detection approaches to measure and model occupancy patterns of the building. As the most widely implemented approach, CO2 concentrations based occupancy acquisition has been used in some studies. However, CO2 concentration based approaches yield to some limitations, such as low sensitivity to large areas, high initial investment for installation, and latency in detection. Developed in recent years, an alternative approach utilizes WiFi technology to automatically detect occupancy information based on existing network infrastructures. However, WiFi signal is unstable and various building components, such as metal separations or concrete walls, can interfere with the signal. In addition, irrelevant network equipment, such as wireless printers or laptops, can misguide the detection probe resulting in the inaccurate occupant count. On the other hand, how to apply WiFi technology in improving occupancy accuracy, occupancy resolution level, or building facility control.

To overcome those research gaps, this thesis conducted three main works. Firstly this thesis utilized the time-series and stochastic characteristics of detected WiFi request to develop a reliable and automatic occupancy detection and prediction mechanism. MAC addresses of devices were used as the identification tags to differentiate occupants and facilities. Several on-site experiments were conducted and the duration time function based WiFi raw data was proposed. This thesis also proposed an occupancy model with Markov chain theory and a further occupancy prediction method with the Markov-based feedback recurrent neural network M-FRNN approach. For comparison, CO2 concentrations based occupancy sensing method was applied while camera based occupancy count was taken as ground truth. The on-site experiment during nine days demonstrated that M-FRNN based occupancy model using WiFi probes showed the best accuracy with a tolerance of 2, 3, and 4 occupants can reach 80.9%, 89.6%, and 93.9%, respectively. As WiFi signal is popularly used in buildings, the occupancy sensing with WiFi signal enables building control systems to adjust services based on people count, which will lead to energy savings and improved occupant comfort. This study provides insights for WiFi-based occupancy studies and occupant-related studies to improve building energy efficiency.

Secondly, this study has investigated the performance of data fusion by combining the environmental parameters data and the WiFi data. To learn and predict the occupancy information, this study selected three feature-based occupancy models using machine learning algorithms of KNN, SVR, and ANN. To assess the occupancy models, the MAE, MAPE, RMSE, and X-accuracy indices were utilized. Results showed that when only the environmental parameters data is applied to learn occupancy, the ANN-based occupancy model was more suitable and accurate, and so would be with the combination of environmental parameters and the WiFi data. The kNN model was more suitable and accurate in learning occupancy information with WiFi data. On the other hand, the three-day accuracy results using WiFi data in occupancy model varied a lot, for example, X-accuracy vary from 81% to 92.1% using kNN, from 76.2% to 89.7% using SVR, and from 75.4% to 81.0% using ANN, but the three-day accuracy results using environmental parameters or both datasets kept in closeness. While MAE, MAPE, and RMSE were not improved a lot when combining environmental parameters and WiFi data. This study proposed an occupancy-profile based multi-zone control strategy to operate a building ventilation system. In this strategy, the WiFi probe technology was implemented to detect real-time and high-resolution occupancy information to determine proper outdoor air portions for all thermal zones of a given building space. To validate the proposed approach, this study conducted an on-site experiment to gather occupancy information for two typical days and used this data simulated the energy and ventilation performance of four different strategies. The results suggested that the proposed the strategy can maintain IAQ in most zones and save up to 55% of energy consumption compared with traditional fixed minimum air flow rate strategy, but it also resulted in unqualified IAQ in 6% to 20.1% of operation time in the critical zone. The proposed method also showed better performance than the ASHRAE recommended the multi-zone equations control strategy.

Finally, the study has illustrated the HVAC load at room level and zone level to provide the foundation of demand-based operation in HVAC systems. Bluetooth Low Energy (BLE) technology was chosen in the occupancy distribution detection as a demand-driven signal. In order to increase the robustness of the algorithm, another feature, location range, was applied, in addition to the received signal strength. The multi features-driven k-NN classification algorithm was utilized to define room occupancy distribution, illustrated by the heat map. In the on-site experiment, six iBeacons were installed and signal fingerprinting methods were applied. To validate the detection accuracy, the experiment outcomes were examined in three case studies, and one City Block Distance (CBD)-based method was used to measure the distance between detected occupancy distribution and ground truth, and to assess the results of occupancy distribution. The results showed the accuracy of CBD = 1 is over 71.4% and the best accuracy of CBD = 2 can reach 92.9%. The results showed that the accuracy of CBD = 0 is about 64.2%. Based on occupancy distribution, we assessed energy performance under the comparisons of actual occupancy information and actual operation with the assistance of wireless sensors nodes. Energy costs to remove thermal load in AO-S-T/RH could be saved 13.81%, 14.16%, 11.69%, and 13.21%, in January, April, July, and October, respectively, and total savings can reach 13.12%. On the other hand, DO-S-T/RH could save 12.04%, 12.3%, 10.11%, and 11.53% compared with energy costs in AO-A-T/RH, in January, April, July, and October, respectively, and total thermal load reduction is 11.4%. This thesis also proposed the idea of controlling HVAC operation in large-scale spaces with high-resolution occupancy distribution. The Dynamic Spatial Occupancy Distribution was defined in this study for interfacing with zone-based VAV operations in large spaces. Under detected occupancy, this thesis proposed one dynamic spatial occupancy distribution-based temperature controller (DTC) to improve HVAC control and compares the proposed DTC with an ideal controller and an existing controller. The results showed about 20.14% 22%of energy conservation in the energy simulation when compared with the conventional controller in two days. The spatial temperature distribution also suggested that the proposed approach was able to maintain occupant thermal comfort without wasting energy by overcooling unoccupied zones. Therefore, the proposed IPS-based demand-driven HVAC control mechanism has been proven to be valid and effective.