Unsupervised Pattern Extraction of Time Series Data for Energy Disaggregation

Today, many information technology applications often produce high-frequency, large-sized, and unlabeled time series data. It is generally laborious and expensive to observe individual components that affect the form of the time series data. However, by clustering the large-scale unlabeled data, it may be possible to detect anomalies, subsequences, events, state changes, and more. The application area of time series clustering is very wide, including analyzing energy load monitoring. With current technologies, we can monitor energy consumption con-tinuously at relatively high frequencies. This makes it possible to disaggregate electrical energy monitoring data obtained only from measurements made at the main breaker level of a house or facility and estimate the individual consumption of devices. There are various studies on disaggregation of load monitoring data, but these studies generally use supervised learning methods. In addition to obtaining appliance consumption separately and detecting anomalies, disaggregated energy load data may be used to determine consumption habits and even detect human activity to some extent. In this study, general information about unsu-pervised segmentation and clustering of time series is given, with the disaggregation of energy load monitoring data mentioned as an application example.