Residential Household Non-Intrusive Load Monitoring Using Multi-Label Classification Methods

Abstract

Smart grid provides a sustainable and environment-friendly vision of the future power systems by incorporating smart meter, remote control and communication technologies into the network of electricity generation, distribution and consumption. Residential Demand-Side Management (RDSM) is an important part of smart grid since a large part of electricity is consumed by residential sector. It refers to programs implemented at the residential customers' side and can be utilized to reduce the overall load demand, most importantly, the peak demand. Since the grid is designed for the peak demand instead of the average demand, the grid is under-utilized for most of the time. By reducing the peak demand or shifting load to off-peak time, the peak-to-average ratio is reduced, grid reliability can be improved, energy can be saved coupled with carbon dioxide and other greenhouse gases emission can be minimized. However, currently it is impossible for residential customers to identify ways to save electricity. On the one hand, residents have no idea of each share of the total electricity consumed by individual appliances since they only receive an aggregate monthly bill. On the other hand, there are more and more electrical appliances in a household nowadays. In order to solve this problem, Non-Intrusive Appliance Load Monitoring (NIALM)was proposed to disaggregate energy consumption to appliance level or circuit level, specifically, it refers to inferring what appliances are operating and how much energy they consuming in a household at a given time solely from house-level aggregate measurements from the main panel. It is one approach to residential demand-side management strategies in the Smart Grid and is now commonly implemented via machine learning. However, by and large, the learning-based NIALM algorithms to date have been single-label approaches; each feature vector is associated with a scalar, categorical value (which means there is only a single appliance associates with an instance of aggregate power). This, however, is a poor match to the NIALM problem, in which multiple independent concepts (active appliances) may simultaneously hold true. To model this characteristic, multilabel classification algorithms (which associate a vector of categorical variables to each feature vector) have been employed for NIALM in this thesis. Furthermore, those learning algorithms generally require a ground-truth classification of the observed fluctuations into the set of appliances then operating; data which is not ordinarily available. As a compromise, we examine semi-supervised learning algorithms, which only need a ground truth from a small sample (e.g from an initial ``registration" period). Thus this thesis presents multi-label algorithms for NIALM, either supervised or semi-supervised, to recognize the operational states of appliances simultaneously, and based on the states information, appliance energy consumption can be calculated. Extensive experiments have been conducted on five public datasets and comparisons have been made against other methods in the-state-of-the-art literature, which prove that graph-based manifold semi-supervised multi-label classification might be a promising approach to NIALM. In the end, a reliance weighting strategy is proposed to improve the performance of graph-based manifold multi-label classification. And extensive experiments have been carried out on another four general multi-label datasets.