DynamoPMU: A Physics Informed Anomaly Detection and Prediction Methodology using non-linear dynamics from $μ$PMU Measurement Data
This addresses the problem of situational awareness in electrical distribution systems for grid operators, though it is incremental as it builds on existing HAVOK-based methods for a specific domain.
The paper tackles the challenge of detecting and predicting infrequent, uncertain events in electrical distribution systems using high-resolution μPMU data by developing DynamoPMU, a physics-informed method based on the HAVOK operator that separates linear dynamics from anomalies. The result is an unsupervised framework demonstrated on real data from the LBNL grid, enabling real-time detection and prediction while adapting to varying statistics.
The expansion in technology and attainability of a large number of sensors has led to a huge amount of real-time streaming data. The real-time data in the electrical distribution system is collected through distribution-level phasor measurement units referred to as $μ$PMU which report high-resolution phasor measurements comprising various event signatures which provide situational awareness and enable a level of visibility into the distribution system. These events are infrequent, unschedule, and uncertain; it is a challenge to scrutinize, detect and predict the occurrence of such events. For electrical distribution systems, it is challenging to explicitly identify evolution functions that describe the complex, non-linear, and non-stationary signature patterns of events. In this paper, we seek to address this problem by developing a physics dynamics-based approach to detect anomalies in the $μ$PMU streaming data and simultaneously predict the events using governing equations. We propose a data-driven approach based on the Hankel alternative view of the Koopman (HAVOK) operator, called DynamoPMU, to analyze the underlying dynamics of the distribution system by representing them in a linear intrinsic space. The key technical idea is that the proposed method separates out the linear dynamical behaviour pattern and intermittent forcing (anomalous events) in sequential data which turns out to be very useful for anomaly detection and simultaneous data prediction. We demonstrate the efficacy of our proposed framework through analysis of real $μ$PMU data taken from the LBNL distribution grid. DynamoPMU is suitable for real-time event detection as well as prediction in an unsupervised way and adapts to varying statistics.