Low-complexity low-rate residential non-intrusive appliance load monitoring

Large-scale smart metering deployments and energy saving targets across the world have ignited renewed interest in residential non-intrusive appliance load monitoring (NALM), that is, disaggregating total household's energy consumption down to individual appliances, using purely analytical tools.;Despite increased research efforts, NALM techniques that can disaggregate power loads at low sampling rates are still not accurate and/or practical enough, requiring substantial customer input and long training periods. In this thesis, we address these challenges via a practical low complexitylow-rate NALM, by proposing two approaches based on a combination of the following machine learning techniques: k-means clustering and Support Vector Machine, exploiting their strengths and addressing their individual weaknesses.;The first proposed supervised approach is a low-complexity method that requires very short training period and is robust to labelling errors. The second, unsupervised approach relies on a database of appliance signatures that we designed using publicly available datasets.;The database compactly represents over 100 appliances using statistical modelling of measured active power. Experimental results on three datasets from US (REDD), Italy and Austria (GREEND) and UK (REFIT), demonstrate the reliability and practicality of the proposed approaches.Large-scale smart metering deployments and energy saving targets across the world have ignited renewed interest in residential non-intrusive appliance load monitoring (NALM), that is, disaggregating total household's energy consumption down to individual appliances, using purely analytical tools.;Despite increased research efforts, NALM techniques that can disaggregate power loads at low sampling rates are still not accurate and/or practical enough, requiring substantial customer input and long training periods. In this thesis, we address these challenges via a practical low complexitylow-rate NALM, by proposing two approaches based on a combination of the following machine learning techniques: k-means clustering and Support Vector Machine, exploiting their strengths and addressing their individual weaknesses.;The first proposed supervised approach is a low-complexity method that requires very short training period and is robust to labelling errors. The second, unsupervised approach relies on a database of appliance signatures that we designed using publicly available datasets.;The database compactly represents over 100 appliances using statistical modelling of measured active power. Experimental results on three datasets from US (REDD), Italy and Austria (GREEND) and UK (REFIT), demonstrate the reliability and practicality of the proposed approaches