Automated Step Change Detection in Real-World PMU Data Addressing Practical Implications

This paper introduces a fully automated, unsupervised technique for detecting step changes in synchrophasor measurements, without prior knowledge or learning data features, without applying the complex task of baselining, parameter tuning. Step changes in synchrophasor measurements occur due to shunt switching, controller set point changes, and etc. With increasing grid oscillations from inverter-based resources (IBRs) step changes are vital for tracking grid response and anticipating IBR controller-related oscillations. This study identifies step changes through a nonorthogonal discrete wavelet transform that relies on smoothed gradient estimation. A non-linear filter based on a multiscale point-wise product of wavelet coefficients is proposed, which takes advantage of the broadband feature of step changes within the space of wavelet coefficients. This filter also eliminates unwanted signal elements, thus decreasing the occurrence of false positives. Finally, through the statistical characterization of multi-scale products, we use an adaptive multiscale dependent threshold for detecting steps amidst real-world measurement noise and grid oscillations. Our studies indicate the effectiveness of our method in detecting step changes in synchrophasor data, supported by showcasing real-world case studies and discussing practical applications. Our method also significantly outperforms the relevant baseline detection methods by measuring Accuracy, F1 Score, and Precision under both ambient (noise) and step change conditions.