Modelling of Electrical Appliance Signatures for Energy Disaggregation

Abstract

The rapid development of technology in the electrical sector within the last 20 years has led to growing electric power needs through the increased number of electrical appliances and automation of tasks. In contrary, reduction of the overall energy consumption as well as efficient energy management are needed, in order to reduce global warming and meet the global climate protection goals. These requirements have led to the recent adoption of smart-meters and smart-grids, as well as to the rise of Non-Intrusive Load Monitoring. Non-Intrusive Load Monitoring aims to extract the energy consumption of individual electrical appliances through disaggregation of the total power consumption as measured by a single smart meter at the inlet of a household. Therefore, Non-Intrusive Load Monitoring is a highly under-determined problem which aims to estimate multiple variables from a single observation, thus is impossible to be solved analytical. In order to find accurate estimates of the unknown variables three fundamentally different approaches, namely deep-learning, pattern matching and single-channel source separation, have been investigated in the literature in order to solve the Non-Intrusive Load Monitoring problem. While Non-Intrusive Load Monitoring has multiple areas of application, including energy reduction through consumer awareness, load scheduling for energy cost optimization or reduction of peak demands, the focus of this thesis is especially on the performance of the disaggregation algorithm, the key part of the Non-Intrusive Load Monitoring architecture. In detail, optimizations are proposed for all three architectures, while the focus lies on deep-learning based approaches. Furthermore, the transferability capability of the deep-learning based approach is investigated and a NILM specific transfer architecture is proposed. The main contribution of the thesis is threefold. First, with Non-Intrusive Load Monitoring being a time-series problem incorporation of temporal information is crucial for accurate modelling of the appliance signatures and the change of signatures over time. Therefore, previously published architectures based on deep-learning have focused on utilizing regression models which intrinsically incorporating temporal information. In this work, the idea of incorporating temporal information is extended especially through modelling temporal patterns of appliances not only in the regression stage, but also in the input feature vector, i.e. by using fractional calculus, feature concatenation or high-frequency double Fourier integral signatures. Additionally, multi variance matching is utilized for Non-Intrusive Load Monitoring in order to have additional degrees of freedom for a pattern matching based solution. Second, with Non-Intrusive Load Monitoring systems expected to operate in realtime as well as being low-cost applications, computational complexity as well as storage limitations must be considered. Therefore, in this thesis an approximation for frequency domain features is presented in order to account for a reduction in computational complexity. Furthermore, investigations of reduced sampling frequencies and their impact on disaggregation performance has been evaluated. Additionally, different elastic matching techniques have been compared in order to account for reduction of training times and utilization of models without trainable parameters. Third, in order to fully utilize Non-Intrusive Load Monitoring techniques accurate transfer models, i.e. models which are trained on one data domain and tested on a different data domain, are needed. In this context it is crucial to transfer time-variant and manufacturer dependent appliance signatures to manufacturer invariant signatures, in order to assure accurate transfer modelling. Therefore, a transfer learning architecture specifically adapted to the needs of Non-Intrusive Load Monitoring is presented. Overall, this thesis contributes to the topic of Non-Intrusive Load Monitoring improving the performance of the disaggregation stage while comparing three fundamentally different approaches for the disaggregation problem.