Phasor measurement units (PMU) are providing more data than ever from the electric power grids for use in various applications such as detection of anomalies and data injection attacks, and monitoring of grid performance. Current methods used for data compression are not specifically designed to leverage the underlying structure of the PMU data, and do not facilitate their interpretation. Because machine learning (ML) algorithms typically operate in a feature space and capture the informative characteristics of the signal, a method that processes PMU data that suits ML functionality can advance grid sensor capabilities.
Researchers at Arizona State University have developed a compression and feature extraction scheme for PMU data. Developed from the emerging area of Graph Signal Processing (GSP), the operation resembles techniques used for image, video, and audio processing. By leveraging the graphical form of PMU signals on a power grid, electric loads are modeled as equivalent circuits and a Graph Fourier Transform (GFT) is applied to extract the eigenvectors of the admittance matrix of the system. The most significant compression is obtained when the system is under quasi-steady-state conditions, wherein lower-graph frequency components are representative of the significant coefficients. Conversely, higher frequencies appear when anomalies occur, providing distinct features that can be analyzed by ML techniques.
• Utility system management
• Electric grid monitoring
• Anomaly detection
Benefits and Advantages
• Compresses historical databases of PMU measurements by two orders of magnitude
• Performs feature extraction which can then be used in many machine learning applications