Occupant behaviour pattern modeling and detection in buildings based on environmental sensing

Gomez Ortega, Jose Luis (2017) Occupant behaviour pattern modeling and detection in buildings based on environmental sensing. Doctoral thesis (PhD), Manchester Metropolitan University.

Preview

Available under License Creative Commons Attribution Non-commercial No Derivatives. Download (4MB) | Preview

Abstract

Occupant presence and behaviour have a signi�cant impact on building energy performance. An occupant present in a building generates pollutants like CO2, odour, heat, which can directly change the indoor environment. Because of this change, the occupant may interact with the building environment to maintain the comfort level, for example, he or she may turn on air conditioning systems. Today's Building Energy Management Systems (BEMS) are usually operated based on a �xed seasonal schedule and maximum design occupancy assumption but fail to capture dynamic information. This is both costly and ine�cient. Recent e�orts on exploitation of environmental sensors and data-driven approaches to monitor occupant behaviour patterns, have shown the potential for dynamically adapt BEMS according to real user needs. Furthermore, this occupant information can also be used for other applications such as home security, healthcare or smart environments. However, most of existing models su�er from inaccuracy and imprecision for occupant state classi�cation, could not adaptively learn from real-time sensor input and they mainly focused on single occupant scenarios only. To address these issues, we present a novel data-driven approach to model occupant behaviour patterns accurately, for both single occupant and multiple occupants with real-time sensor information. The contributions can be summarised as follows: Firstly, we have conducted a thorough benchmark evaluation of classi�cation performance of state-of-the-art Machine Learning (ML) methods and occupant related publicly available datasets. Secondly, based on the �ndings in literature and our own experimental evaluations, we have developed a novel dynamic hidden semi-Markov model (DHSMM), which can accurately detect occupant behaviour patterns from sensor data streams in real-time. Thirdly, built upon the online DHSMM model, we have developed a novel incremental learning approach to allow dynamically learning over streaming data. Finally, we have conducted an experimental evaluation of our proposed model Online DHSMM Multi-Occupant for occupancy detection for both single and multiple occupants. We have validated our approach using real datasets and the experimental results show our proposed approach outperforms existing methods in terms of classi�cation accuracy and processing time/scalability. To the best of our knowledge, we have �rst developed a HSMM-based incremental online learning approach to fast and accurate learn building occupant patterns over streaming data for both single and multiple occupants in a holistic way. Additionally, our approach signi�cantly improves the classi�cation accuracies of traditional Markov models (over 10% accuracy increase, while maintaining the model complexity and performing multioccupant detection).