Towards energy-proportional anomaly detection in the smart grid

Phasor Measurement Unit (PMU) deployment is increasing throughout national power grids in an effort to improve operator situational awareness of rapid oscillations and other fluctuations that could indicate a future disruption of service. However, the quantity of data produced by PMU deployment makes real-time analysis extremely challenging, causing grid designers to invest in large centralized analysis systems that consume significant amounts of energy. In this paper, we argue for a more energy-proportional approach to anomaly detection, and advocate for a decentralized, heterogeneous architecture to keep computational load at acceptable levels for lower-energy chipsets. Our results demonstrate how anomalies can be detected at real-time speeds using single board computers for on-line analysis, and in minutes when running off-line historical analysis using a multicore server running Apache Spark