Contributions to Electrical Energy Disaggregation in a Smart Home

Authors

Abstract

Tomorrow's main energy resource may well be energy efficiency. Over the last years, society awareness on environmental changes and high energy costs has been increasing. Nevertheless, the improper use of electrical devices still represents a substantial slice of the electrical energy consumption. Continuous and detailed electricity monitoring has been demonstrated an essential tool to ensure energy efficient in buildings as our homes. Appliance-specific consumption information empowers consumers, leading to informed choices and change of behaviours. Non-intrusive Load Monitoring (NILM) systems, aiming at energy monitoring, load forecasting and improved control of residential appliances, are an attractive solution to bring detailed consumption at device-level to end-users. Using only the aggregated electricity consumption data acquired at a single-point, usually the utility-customer interface, NILM discerns appliances' power usage data employing machine learning and pattern recognition algorithms. Due to its possible low cost, easy installation and easy integration into the future smart grids, which would enable consumers to participate in the electricity market, NILM has become an active area of research. This Thesis is concerned with energy disaggregation as the key part of a NILM framework. Given the whole-home electrical consumption data it aims at investigating and exploring methodologies not yet applied to tackle the correct disaggregation of this signal into the detailed usage of each appliance, or groups of devices, connected to the home electrical circuit. Widespread NILM approaches usually explore the disaggregation of single-point acquired data as a classification problem for which appliances signatures are required. Yet, no set of distinctive characteristics able to accurately describe each appliance has been found. Thereby, this thesis reinforces the search for the set of features used as appliances signatures. Namely, a rule for steady-state identification and its mathematical proof are introduced. This rule was applied for detection of step-changes occurring in the active and reactive power signals and the power factor measurements. The step-changes identified comprised a new appliance signature posteriorly used by the 5-Nearest Neighbours and the Support Vector Machines classification methods in order to obtain the appliance identification. The computational experiments yielded in real-world dataset showed the effectiveness of the proposed signature for distinguishing the different loads in study. The disaggregation and extraction of meaningful information from the aggregated electricity consumption can alternatively be interpreted in the light of signal processing analysis. In this sense, signal processing and time series analysis strategies arise as suitable tools for the extraction of information from the whole-home signal. Before aiming at the calculation of consumption estimates for each appliance, a previous study concerning the extraction of variations in the aggregated electrical signal associated with devices that work automatically without any human intervention is performed. In this context, a technique based on Wavelet Shrinkage and signal processing operations, designed to extract information from the aggregated signal considering several of its segments that can be analysed by distinct mother wavelets, is proposed. Following this path, a novel way to look into the issue of energy disaggregation is its interpretation as a single-channel source separation problem. To this end, the performance of source modelling based on multi-way arrays (tensors) and correspondent factorization is analysed. With the proviso that a tensor composed by the data for the several devices in the house is given, non-negative tensor factorization is performed in order to extract the most relevant components. The outcome is later embedded in the test step, where only the whole-home measured consumption is available. Inference of individual consumptions is then achieved by matrix factorization using the learned models. The approaches based on signal processing, for the extraction and disaggregation of information from the whole-home electrical signal, were successfully evaluated on a real-world dataset, as illustrated by the favourable performance and statistical evidence. Overall, this Thesis contributes with electrical energy disaggregation approaches, successfully validated on real-world data, which - as we hope - will have a positive impact in solving efficiency problems in a smart home.

Keywords

Eletrical Energy Disaggregation, Single-Channel Source Separation, Feature Extraction and Pattern Classification

Subject

Non-Intrusive Load Monitoring Systems, Energy Disaggregation

PhD Thesis

Contributions to Electrical Energy Disaggregation in a Smart Home 2013

PDF File