Exploring PMU Measurement Representation for ML-based Fault Diagnosis Models

Synchrophasors are becoming more available to distribution systems through the use of Phasor Measurement Units (PMU). This enables the performance of Machine Learning (ML) based fault diagnostic models to improve due to utilization of underlying information from the angle measurement instead of only using the magnitude measurements. However, most ML models do not support complex numbers, such as phasors, as input. One strategy is to decompose each phasor measurement into two features which can be represented in either rectangular or polar form. This research provides a basis on determining which representation yields better patterns in the feature space in terms of building better performing fault diagnosis models. The patterns in the feature space due to different representations were observed through visual and empirical evidences. Results show that decomposing the phasor into its rectangular representation yields better patterns in the feature space which could result in more accurate and robust fault diagnosis models.