Unsupervised Learning for Distribution Grid Line Outage and Electricity Theft Identification

The development of smart meters enables situational awareness in electric power distribution systems. The situational awareness provides significant advantages such as line outage and electricity theft detection. This paper aims at using smart meter data to detect these anomalies. To do so, an appropriate cluster-based method as an unsupervised machine learning approach is applied. A stochastic method based on conditional correlation is also proposed to localize the anomalies. It is shown that this can be done by detecting changes in bus connections using present and historical smart meter data. Therefore, network topology inspection can be avoided if the proposed method is applied. A complex mesh grid is used to demonstrate performance of the data-driven anomaly detection approach. The results show that determining the right value of hyper-parameter and adequate features extraction leads to acceptable accuracy.