Hierarchical Temporal Memory Based Machine Learning for Real-Time, Unsupervised Anomaly Detection in Smart Grid: WiP Abstract

Micro-phasor measurement unit (μPMU) sensors in smart electric grids provide measurements of voltage and current at microsecond timescale across the network and have great potential value for grid diagnostics. In this work, we propose a novel neuro-cognitive inspired architecture based on Hierarchical Temporal Memory (HTM) for real-time anomaly detection in smart grid using μPMU data. The key technical idea is that the HTM learns a sparse distributed temporal representation of sequential data that turns out to be very useful for anomaly detection in real-time.Our numerical results show that the proposed HTM architecture can predict anomalies with 96%, 96%, and 98% accuracy for three different application profiles namely, Standard, Reward Few False Positive, Reward Few False Negative, respectively. The performance is compared with three state-of-the-art real-time anomaly detection algorithms and HTM demonstrates competitive score for real-time anomaly detection in μPMU data.