A KDE-based Methodology for real-time PMU Data Management and Oscillation Detection

The increase in renewable penetration in power grids has led to more frequent system oscillation occurrences. Some, if left unattended, will deteriorate to oscillatory instability that causes severe damages. While fast power system variations can be captured by Phasor Measurement Units (PMU), critical oscillations are often blended with common grid dynamics and a data management method is necessary to differentiate the two. Currently, the widely-adopted real-time oscillation event detection methods are mostly empirical-based, which present a large number of false alarms and omissions even with constant tuning and tweaking. Therefore, to enhance notification accuracy and facilitate control room functions, a real-time oscillation detection methodology based on Kernel Density Estimation (KDE) was innovated and deployed in AEP. The proposed methodology fits millions of historical PMU data into a bi-variate probability model by superposing Gaussian kernel functions derived at each data point. The standard deviations of those Gaussian kernel functions are determined via Maximum Likelihood Cross Validation (MLCV), in which the optimum fitness between the probability model and the historical PMU data are achieved. A severity index is then created for each observation based on the cumulative probability function. A cutoff value on the severity index is subsequently customized for every PMU signal via heuristic training involving historical event data. Outliers, picked based on the cutoff severity index, would hence be visualized on scatter plots to delineate alarm thresholds and finalize alarm settings. The KDE-based oscillation detection was able to flag critical oscillations previously overlooked by the empirical-based alarm. Moreover, it has decreased false notification rate from 50% to below 6% since deployment in 2020. Despite of evolving changes in AEP’s footprints, the false alarm rate has been maintained around 5%, which demonstrates the robustness of the proposed methodology. The KDE-based oscillation detection allows control room personnel to capture threatening oscillations at an early stage, enhancing situational awareness and preventing system-wide instability. Furthermore, with post-event studies performed on captured events, preventive measures could be planned accordingly to accommodate today’s power grid to increasing renewable penetrations. Co-sponsored by: University of Tennessee, Knoxville Speaker(s): Yidan Lu Virtual: https://events.vtools.ieee.org/m/360929