Time-series decomposition of power demand data to extract uncertain features

The spread of smart meters means that a large amount of power demand information from private houses is being collected around the world. Owing to the development of smart city infrastructure, the use of standardized frameworks for extracting features from power demand information has become vital. In this paper, we propose a novel decomposition approach useful for extracting feature values from power demand information from a house. Energy consumption was monitored for multiple houses for one month in Japan with a sampling duration of 30 minutes, which is a standard sampling time of smart meters in Japan. First, periodic characteristics were detected for 24 hours based on autocorrelation analysis. Then, the monitored information was decomposed into four components: standby power, trends, and periodic and residual parts. The distribution of the residual part is similar to a Gaussian distribution, so the behavior of the residual part was parameterized using variance and average. Trend, periodic, and residual components were clustered by means of k-means clustering in order to aggregate the difference in behaviors. There was no periodic component in the residual part according to auto-correlation analysis. Nevertheless, some clusters had a relatively large variance, which means that abnormal power demand occurred frequently in datasets. The amount of variance and climate correlation was analyzed, and the fact detected that large scale events disturb usual daily life-styles, from the viewpoint of energy usage. Last, these features were compared with actual customer information. In the evaluation, family structure and floor space were utilized to prove the effectiveness of the proposed decomposition approach. The evaluation proved that this decomposition method could extract uncertainty features from power demand information.