|Abstract: ||Many countries are rolling out smart electricity meters. A smart meter measures the aggregate energy consumption of an entire building. However, appliance-by-appliance energy consumption information may be more valuable than aggregate data for a variety of uses including reducing energy demand and improving load forecasting for the electricity grid. Electricity disaggregation algorithms – the focus of this thesis – estimate appliance-by-appliance electricity demand from aggregate electricity demand.
This thesis has three main goals: 1) to critically evaluate the benefits of energy disaggregation; 2) to develop tools to enable rigorous disaggregation research; 3) to advance the state of the art in disaggregation algorithms.
The first part of this thesis explores whether disaggregated energy feedback helps domestic users to reduce energy consumption; and discusses threats to the NILM. Evidence is collected, summarised and aggregated by means of a critical, systematic review of the literature. Multiple uses for disaggregated data are discussed. Our review finds no robust evidence to support the hypothesis that current forms of disaggregated energy feedback are more effective than aggregate energy feedback at reducing energy consumption in the general population. But the absence of evidence does not necessarily imply the absence of any beneficial effect of disaggregated feedback. The review ends with a discussion of ways in which the effectiveness of disaggregated feedback may be increased and a discussion of opportunities for new research into the effectiveness of disaggregated feedback. We conclude that more social science research into the effects of disaggregated energy feedback is required. This motivates the remainder of the thesis: to enable cost-effective research into the effects of disaggregated feedback, we work towards developing robust NILM algorithms and software.
The third part of this thesis describes our effort to advance the state of the art in disaggregation algorithms. Three disaggregation approaches based on deep learning are discussed: 1) a form of recurrent neural network called ‘long short-term memory’ (LSTM); 2) denoising autoencoders; and 3) a neural network which regresses the start time, end time and average power demand of each appliance activation. The disaggregation performance was measured using seven metrics and compared to two ‘benchmark’ algorithms from NILMTK: combinatorial optimisation and factorial hidden Markov models. To explore how well the algorithms generalise to unseen houses,the performance of the algorithms was measured in two separate scenarios: one using test data from a house not seen during training and a second scenario using test data from houses which were seen during training. All three neural nets achieve better F1 scores (averaged over all five appliances) than either benchmark algorithm. The neural net algorithms also generalise well to unseen houses.|