Uma abordagem baseada em árvores de decisão para a análise da estabilidade angular do rotor

Abstract

The power system security assessment is essential to ensure the supply of electrical energy and the feasibility of the operation. Among these analyses, the study of the rotor angle stability aims to ensure that the electromechanical modes of the system are well damped and, for that, it is common that the methods employed make use of mathematical models of the electric power system. However, the dependence on mathematical models, potentially, makes the evaluation methods sensitive to variations in parameters and changes in the network that have not been adequately represented, mainly in realtime applications. Currently, Smart Grids proposes to offer greater monitoring capacity and sample rates that allow real-time analysis. Concomitantly, the advances in distributed and cloud computing have encouraged the use of machine learning techniques to solve various problems using the massive amount of data available. In this sense, this work proposes to make use of the measurements made available by phasor measurement units to evaluate the feasibility of using decision trees in the analysis of the rotor angle stability. For stability analysis at small disturbances, a decentralized approach based on individual decision trees and data from phasor measurement units allocated in the generator buses is applied. In this approach, each decision tree uses only local measurements to evaluate the rotor angle small-signal stability, in this way, the classification can be carried out even when there is loss of information from specific generators or failure in the communication system. When the system is subjected to a large disturbance, a second method is employed, which is based on a centralized decision tree and voltage phasors measured at generator buses from the whole studied system. This last classifier is able to identify instability in the response post fault portion of the system, establishing a trade-off between the number of measurement cycles used and the classifier performance. The results obtained on the IEEE 68-buses system showed the efficiency of the proposed approach. In the classification of small-signal stability, an accuracy of 93% is reached by the distributed trees even in scenarios with contingencies and load variations. About large disturbances was possible to classify with precision the transient stability even with only 1 measurement cycle (96.4%) and with only 3 measurement points along the test system.