Extracting Discriminative Features for Event-based Electricity Disaggregation

We describe a novel method for electricity load disaggregation based on the machine learning method of time series shapelets. We frame the electricity disaggregation problem as that of event detection and event classification from time series data. We use existing shapelet-based algorithms to separate appliance activity periods (caused by switching on/off of appliances and denoted as events) from time periods without any such activity. We then identify which type of appliances in a household correspond to the events detected within the power consumption data. Such appliance-level feedback is critical for end-users in managing their energy use efficiently. We use the BLUED dataset for experimental evaluation of the proposed method. This dataset is a fully labeled publicly available dataset of electricity consumption of a household in the United States for one week, the data being recorded at a very high frequency and externally labeled with the times when specific appliances were switched on or off. The proposed approach is able to achieve approximately 98% accuracy for event detection and between 77% to 84% accuracy for event classification. The data segments that were identified as being most discriminative for electricity disaggregation are visually interpretable, and the appliances identified to be responsible for heavy energy consumption can be reported to consumers to encourage reduction in energy consumption.