Abstract

Smart buildings use intelligent automation systems which optimize energy consumption while improving occupant comfort and promoting sustainable development. The core of this vision de-
pends on strong Occupancy Estimation and Activity Recognition models which support dynamic control of HVAC systems and lighting and other essential building operations. These models face
significant deployment challenges because real-world settings differ from training environments and suffer from insufficient availability of labeled data and evolving activity patterns. This thesis examines how Open Set Domain Adaptation, Universal Domain Adaptation, and Generalized Domain Adaptation enhance the adaptability and generalization potential of Occupancy Estimation and Activity Recognition models in smart buildings. Our research begins with the exploration of Open Set Domain Adaptation techniques designed to distinguish known activity classes from unknown
ones during domain shifts by implementing adversarial learning frameworks along with rejection-aware classifiers specifically for smart building sensor data. Our work presents a combined Uni-
versal Domain Adaptation approach which uses optimal transport and angular margin constraints to achieve flexible alignment between domains while operating without knowledge of overlapping
labels. Our study examines generalized domain adaptation methods that allow adaptation across domain and label shifts in previously unencountered environments through self-training and hybrid learning approaches combined with distribution-agnostic strategies. Extensive experiments show that the proposed methods excel in classification accuracy, unknown class detection capabilities and stability against label imbalance. The research provides scalable, privacy-conscious solutions for adaptive behavior modeling in intelligent environments, which help improve the energy efficiency of intelligent building systems.