Real-time multi-state classification of short-term voltage stability based on multivariate time series machine learning

Abstract

The analysis of the significant amount of data collected by PMU devices in wide-area monitoring, protection and control (WAMPAC) applications has been a great challenge for assessing power system fast phenomena. This paper presents a novel methodology for real-time assessment of short-term voltage stability (STVS) under large disturbances with an approach based on data mining and machine learning. This methodology classifies off-line the power system stability in multiple operating states through the calculation of the maximal Lyapunov exponent and dynamic voltage indices. This allows identifying not only fast voltage collapses (FVC) but also fault-induced delayed voltage recovery (FIDVR) events. The multivariate time series data of the power system dynamic response are processed and transformed with a symbolic representation technique. This together with the operating state classification are used to train an intelligent machine based on Random Forest proposed for applying in real time to classify the post-disturbance operating state. This methodology is tested in the New England 39-bus system. The performance of the methodology to classify the STVS in real time was verified, obtaining a classification error less than 2% using a post-contingency data window of 0.58 s and less than one-third of all the bus voltage measurements. Results show the ability of the methodology to predict voltage stability problems, having enough time to carry out automatic control actions to prevent or mitigate problems.