Big Data Platform for Real-Time Oscillatory Stability Predictive Assessment Using Recurrent Neural Networks and WAProtector's Records

Abstract

After a perturbation, the generators shift their operating condition in search of new equilibrium states (steady states), overpassing a dynamic .state (which should be transitory), characterized by power and frequency oscillations. Oscillations are marked by the so-called oscillation modes that are determined by three fundamental parameters: Amplitude (MW), Frequency (Hz), and Damping Ratio (%). These oscillatory modes can be estimated in real-time using modal estimation algorithms applied to signals recorded by Phasor Measurement Units (PMUs) within a Wide Area Monitoring System (WAMS). These estimations are made each time a new sample arrives, so they do not provide predictions of the future status of oscillatory stability. However, an aspect of relevance in the operation of electric power systems is the need for the operator to have "early warnings" that allow him to make decisions sufficiently in advance to carry out control actions. In this sense, it is necessary to have short-term prediction mechanisms (a few seconds in the future) of the modal analysis results, which allow the operator to anticipate the evolution of the operating state to predictively evaluate the oscillatory stability of the system. In this sense, a Big Data platform to analyze the streaming data that comes from WAMS, being capable of analyzing the data from the modal estimation and performing a predictive evaluation, automatically, of the oscillatory stability status, is proposed. Therefore, this work presents the platform's key implementation aspects, which are based on Data Management Technologies (Cassandra), together with a Data Analytics software (Python), in which a time series regressor is trained based on recurrent neural networks (RNN). This methodology is applied to the Ecuadorian Electric Power System, taking advantage of its WAMS platform WAProtector.