Metodologia de classificação de eventos em redes de distribuição ativa de energia utilizando curvas V-I

Abstract

The operation of Electrical Power Systems has undergone significant transformations since the introduction of renewable energies into the global energy matrix. Specifically in distribution systems, the utilization of Distributed Energy Resources has brought the concept of Active Distribution Systems to the forefront. The aspect of descentralized generation in these systems gave way to a complexity before unseen in conventional distribution, which are characterized by centralized generation, in terms of monitoring and protection of the electrical system. This way, the introduction of the Phasor Measurement Unit (PMU), which utilizes synchrophasor technology, for active grids monitoring gained space in place of conventional Supervisory Control and Data Acquisition systems (SCADA) for having a larger capacity of electrical grid data storage in large quantities. Through the use of mathematical techniques, like the Discrete Fourier Transform (DFT), for phasor estimation, PMUs are able to calculate phasors of harmonic components of voltage and current signals in the system, acquiring important signal features of complex events, like High Impedance Faults. In this context, methods that use machine learning models for active systems monitoring became popular, as these models’ capacity for detection and classification of varied classes of events through data processing and feature extraction for training. Therefore, this work develops an event classification methodology in an active distribution system using a machine learning model. First, 7 MATLAB datasets composed by data of 3 events, Low Impedance Faults, High Impedance Faults and Capacitor Bank Switching, were built, obtained from buses of an active distribution benchmark system proposed by the International Council on Large Electric Systems (CIGRE). These datasets consist of voltage x current (V-C) curves obtained from PMU readings of the original signals time series plus V-C curves consisting of harmonic phasor modules from 1°to 7°order (calculated from these same original signals) of these events. 1 of the datasets was built with unaltered data obtained from simulations. For evaluation of the model’s robustness, 3 datasets were built simulating applied noise, relative to the voltage and current signals, of 60 dB, 40 dB and 20 dB and 3 other datasets were built simulating data loss in these signals of 5%, 10% and 30%. After building these datasets, they were applied in three bidimensional multiclass classification Convolutional Neural Networks, all of which use the same architecture, in Python language, utilizing the TensorFlow and Keras libraries. One model classifies the events with the original signals and 1°to 7°harmonic V-C curves. The other two models, in this same range, use these signals’ characteristics only with pair and odd harmonics, respectively. All three models obtained satisfactory test results, obtaining average accuracies, respectively, of 99,05%, 98,81% and 99,34%. Such results demonstrate that the trained models are capable of classifying the events in the presented active system eficiently.