A Deep Unsupervised Learning Approach to PMU Event Detection in an Active Distribution Network

The increased deployment of advanced metering devices such as PMUs in active distribution networks offers network operators the potential to develop applications which may help improve visibility regarding the operating state of their network. Considering the high reporting rates of PMUs, it is imperative to develop schemes capable of analysing PMU data in near real-time, which will enable system operators take timely actions. However, current methods used for situational awareness in active distribution grids are heavily dependent on data sample size or on detectors that require great effort in tuning. By training deep autoencoders with PMU data from a real active distribution network, we develop and validate a framework which automatically learns a representation of the network. With this approach, our framework can monitor the state of the network and also detect events in a fraction of a second, making it suitable for online analysis of PMU data from multiple streams or analysis of large amounts of historical PMU data. The performance of the proposed model is validated by comparison with a sample based statistical anomaly detector.