Improving Accuracy of Occupancy Detection by Estimating Actual Demand from Smart Meter Data

Abstract

In this paper, we propose high-accuracy occupancy detection using low-resolution electricity consumption data. In Japan, residential smart meters, which automatically read and transmit energy consumption data at each household to electric power companies, have started to be installed and will be set up in 80 percent of households by 2020. Occupancy detection is one of the major techniques leveraging electricity consumption data and is applicable various services such as ambient assisted living and peak load shifting. However, it is difficult to conduct high-accurate occupancy detection using the transmitted smart meter data because they are in 30-minute interval and truncated to 100Wh units. Especially, truncation makes difficult to analyze the change of demand by absence. Therefore, we propose a machine-learning based occupancy detection method combined with the estimation of the actual consumption from the truncated data using total variation regularization. In experiments, our method shows the performance is comparable to the result using the raw demand data in 1W unit.