Abstract
Technological advancements in sensing, networking and computation opened up the possibilities to sense user-centric information to solve many problems such as conservation of energy in commercial buildings. Research on leveraging such capabilities to optimize the energy utilization in a facility or a building is relatively new. The current thesis presents a framework that capitalize on heterogeneous sensing infrastructure present in a smart office space to track operational states of the appliances without the need to deploy energy meter on every device of interest. This study extends techniques from Non-intrusive Load Monitoring (NILM) domain that automates detection of operational appliance activities using aggregated load measurements, by employing sophisticated signal processing and machine learning algorithms. This study also addresses challenges such as the inability of existing methods to accurately localize and characterize state transition events of low-power appliances due to similarity of their power consumption profile. In addition, this study demonstrates how the effectiveness of traditional approaches has been compromised by their ability to recognizing multistate appliance operations due to lack of robust appliance signatures extracted from low-granularity power measurements. As a result, this study explores event detection and characterization mechanisms that includes the application of singular spectrum transformation for improved event localization, and extraction of new features to enhance class discrimination between target appliances. In addition, it proposes a multi-modal event characterization framework to deal with appliance classes that exhibit ambiguous overlap of power signatures in a feature space. The aim is to create a unified hybrid space by characterizing the power and acoustic profile of appliances and optimally combine them using kernel-based feature fusion strategy. The study demonstrates how the proposed system can better distinguish between appliances of different categories in this new feature space and consequently achieves a higher appliance state estimation accuracy. To evaluate the suitability of non event-based models for load disaggregation, a specialized variant of the hidden Markov model (HMM) known as factorial HMM is investigated for inferring appliance states based on aggregated load measurements. To demonstrate this approach in the real world, a mobile phone application was developed and evaluated in actual practice. In addition to load disaggregation, an interrelated challenge is to identify abnormal or unusual consumption patterns within specific energy measurements. Due to the high volume and noise content of sensor readings, data compression and appropriate feature representations were essential for effective analysis of energy measurements. To address these challenges, this study proposes an anomalous load pattern detection framework that performs wavelet approximation of electrical load curves, and further reduces their dimensionality using the classical multidimensional scaling method (CMDS). Results showed that the low-dimensional projection of features prior to performing anomaly detection effectively isolate the anomalous patterns and as a result improves the performance of target anomaly detection models.