Two unsupervised learning algorithms for detecting abnormal inactivity within a household based on smart meter data

Abstract

We study the problem of detecting abnormal inactivities within a single-occupied household based on smart meter readings. Such abnormal events include immobilizing medical conditions or sudden deaths of elderly or disabled occupants who live alone, the delayed discovery of which poses realistic social concerns as the population ages. Two novel unsupervised learning algorithms are developed and compared: one is based on nested dynamic time warping (DTW) distances and the other based on Mahalanobis distance with problem-specific features. Both algorithms are able to cold-start from limited historical data and perform well without extended parameter tuning. In addition, the algorithms are small profile in terms of data usage and computational need, and thus are suitable for implementation on smart meter hardware. The proposed methods have been thoroughly validated against real data sets with simulated target scenarios and have exhibited satisfactory performance. An implementation scheme on smart meter hardware is also discussed.