Learning-based Real-time Outage Location Identification in Power Distribution Systems with Sparse Sensors

Real-time identification of outage locations in power distribution systems is an essential task for utilities to mitigate the negative impact of power system outages. In this paper, a learning-based model-assisted method is developed for identifying outage locations in power distribution systems in real time. Notably, the method utilizes primarily only a) sparsely located supervisory control and data acquisition (SCADA) measurements, and potentially b) last gasp signals from a small number of smart meters. The method exploits offline training of outage location predictors based on data simulated with synthetically generated load profiles and outage scenarios. The trained predictors can then be used in an online fashion to accurately identify outage locations in new scenarios in real time. With physical-model-based network partitioning, the offline learning is decoupled into training predictors for much smaller sub-regions so that the learning efficiency is much improved. Importantly, the trained predictors based only on SCADA measurements and entropy loss functions can be integrated with smart meter last gasp signals, without toss of any optimality, regardless of smart meter locations, outage scenarios, and performance evaluation metrics. Evaluation of the method based on real-world power distribution feeder and load data demonstrates high accuracy in outage location identification even using SCADA measurements only. We then demonstrated how having just a handful of smart meters with last gasp capabilities can further improve the outage location accuracy significantly.