Shape-based data analysis for event classification in power systems

Fault classification in power systems is a challenging and complex task as the variety and variability of the electrical parameters of the various network components in spatial and temporal scales. The majority of machine learning methods for event detection require the labeled data sets or examples of previous events. However, the recorded event data happen in different locations, time and system conditions. Therefore, they are not aligned time-series which introduce more challenges for feature selection and signal processing. To perform better feature selection for time-series measurements, shape-based methods along with time alignment (also called registration) are needed. This paper presents the Fisher-Rao Registration Method (FRMM) as a solution for the alignment of different time signals. Amplitude and phase components resulting from the Fisher-Rao registration method provide a means for implementing a hierarchical clustering analysis classifying different fault events by type. The algorithm was tested with the IEEE 13-nodes test feeder simulated in RSCAD environment with over 1500 different fault events presenting an average prediction rate of 98%.