Voltage Sag Source Classification using Multivariate Time Series and Soft Dynamic Time Warping

Voltage sags are a major concern due to their frequency and impact on electrical equipment. Determining their underlying cause and relative location is a first step towards the application of adapted mitigation solutions. In this paper we propose a methodology for the classification of voltage sag sources based on multivariate time series signatures obtained through the application of Short-Time Fourier transform and Fortescue transform. These signatures or patterns are then classified following a "nearest neighborhood" approach, by reducing clusters of time series into a single multivariate time series centroid per class using the soft-Dynamic Time Warping algorithm. This approach overcomes the limitations of the 1-Nearest Neighbors with Dynamic Time Warping algorithm in terms of robustness regarding outliers and computational cost. Moreover, we evaluate the generalization capabilities of the algorithm using combinations of synthetic and real field data coming from three industrial sites. We obtain a simple, interpretable and easy-to-implement classifier capable of using 100% synthetic data for database and achieving 99.74% of accuracy on real data.