Non-parametric modeling in non-intrusive load monitoring

Non-intrusive Load Monitoring (NILM) is an approach to the increasingly important task of residential energy analytics. Transparency of energy resources and consumption habits presents opportunities and benefits at all ends of the energy supply-chain, including the end-user. At present, there is no feasible infrastructure available to monitor individual appliances at a large scale. The goal of NILM is to provide appliance monitoring using only the available aggregate data, side-stepping the need for expensive and intrusive monitoring equipment. The present work showcases two self-contained, fully unsupervised NILM solutions: the first featuring non-parametric mixture models, and the second featuring non-parametric factorial Hidden Markov Models with explicit duration distributions. The present implementation makes use of traditional and novel constraints during inference, showing marked improvement in disaggregation accuracy with very little effect on computational cost, relative to the motivating work. To constitute a complete unsupervised solution, labels are applied to the inferred components using a Res-Net-based deep learning architecture. Although this preliminary approach to labelling proves less than satisfactory, it is well-founded and several opportunities for improvement are discussed. Both methods, along with the labelling network, make use of block-filtered data: a steady-state representation that removes transient behaviour and signal noise. A novel filter to achieve this steady-state representation that is both fast and reliable is developed and discussed at length. Finally, an approach to monitor the aggregate for novel events during deployment is developed under the framework of Bayesian surprise. The same non-parametric modelling can be leveraged to examine how the predictive and transitional distributions change given new windows of observations. This framework is also shown to have potential elsewhere, such as in regularizing models against over-fitting, which is an important problem in existing supervised NILM.