User-centric smart sensing for non-intrusive electricity consumption disaggregation in buildings

2017-01-10This dissertation focuses on enabling energy aware facilities through the application of low cost advanced sensing systems using non‐intrusive load monitoring (NILM) for buildings. Acquiring disaggregated electricity consumption information at individual load level enables energy aware facilities and provides the ground for various applications, such as human energy‐affecting behavior modification, load shedding and demand‐response applications, fault detection, activity detection, etc. As a low‐cost alternative to appliance level sensing, NILM uses few sensing nodes for measuring aggregated electricity data, coupled with specialized machine‐learning and signal processing algorithms to infer operational schedules of individual loads. Majority of the NILM solutions are based on supervised learning algorithms and thus require training data, for which user interaction with the NILM system is required due to diversity of appliances’ technologies and signatures. The need for training is one of the obstacles in wide commercial adoption of NILM systems. Accordingly, this research focuses on user‐centric intelligent sensing system, in which NILM is defined as a problem of human‐computer interaction in the context of load disaggregation. The main goal of the research is to facilitate the adoption of NILM systems by considering two objectives of relaxing user‐interaction requirements and time constraints during the training period while preserving accuracy. This dissertation has contributed by devising a framework for smart interaction in the form of a user‐centric NILM system, through which clusters of appliances’ signatures are autonomously populated for intelligent communication (versus a brute‐force approach) with users. The objective was to enable clustering of the appliances transient signatures (represented in a multiscale high dimensional feature space) compatible to the appliances’ operational modes in the physical domain. Accordingly, a customized cluster validation approach, which uses a combination of cluster splitting and merging, was developed. As part of this approach, an unsupervised recursive hierarchical clustering based heuristic algorithm was devised and comprehensively evaluated. This autonomous clustering algorithm was coupled with unsupervised cluster merging and internal validation to enable accurate feature space partitioning. The autonomous clustering module was integrated with data acquisition and processing, event detection, classifier, and anomaly detector components to achieve efficient interactions while keeping high level of accuracy. An ensemble of anomaly detection algorithms combination was introduced to balance the trade‐off between inlier detection rate and accuracy metrics. The framework enables efficient interactions through reduced user‐system communication for labeling populated clusters of data. The evaluation of the smart interaction framework showed more than 94% reduction in the number of interactions, while achieving the labeling accuracy more than 96% on average. Moreover, we have coupled this framework with a “passive” training algorithm, which enables users to interact with NILM system under relaxed time constraints instead of reacting to the operational schedule of appliances. A signature‐label matching algorithm was presented to enable the NILM system to match signatures with corresponding labels with a user reaction time relaxation of about 30 minutes. These algorithms have been validated in three residential settings through implementation of a user centric real‐time prototype with a single sensing point at the main circuit panel. The research has also contributed to the field by being the first on quantification of the training interaction and also by providing fully labeled (acquired in high frequency) data sets in three real residential settings, where user interaction with the NILM system for training has been recorded