Predição da estabilidade a pequenas perturbações em sistemas de energia elétrica: uma abordagem considerando as incertezas da geração eólica

Abstract

In the current context, where the demand for sustainable energy sources with low environmental impact is increasing, wind power generation plays a prominent role. However, the high penetration of this energy source imposes new challenges to the operation of Power Systems, particularly involving small-signal stability and the behavior of interarea electromechanical oscillation modes. These challenges become even more critical due to the need for real-time assessment of these modes, which can be supported by Phasor Measurement Units and analytical techniques capable of handling the variability and uncertainty inherent in renewable generation. In this context, this work aims to develop a methodology for small-signal stability prediction in systems with high wind power penetration, using machine learning techniques with a focus on decision trees. This approach enables stability prediction at the pre-fault instant, with low computational cost and fast processing, favoring real-time application. To achieve this, the IEEE 68-bus system was modified by incorporating approximately 40\% of wind power generation and considering variations in load, wind speed, and contingencies according to the $N-1$ criterion, thereby composing the study database. Modal characteristics were then extracted using the Prony method. Next, two approaches for critical mode classification were evaluated: a centralized approach, which aggregates data from multiple PMUs into a single predictive model, and a decentralized approach, based on local models. In both approaches, classifications were analyzed in both binary (stable or unstable) and three classes form (secure, insecure, or unstable), in order to investigate not only binary stability prediction but also the capability of providing more detailed information to the system operator. The centralized approach achieved the best performance, particularly in the three-class classification, reaching an accuracy above 96\% and without errors at critical classification boundaries, meaning without misclassifying unstable cases as stable and vice versa. The decentralized approach, although showing lower accuracy in the three-class classification, presented satisfying results in binary classification, with accuracy above 99\%, proving suitable for contexts with communication constraints or partial system failures. The results confirm the effectiveness of the proposed methodology for small signal stability prediction in systems with high wind power generation penetration, providing support for safe operation and informed decision-making under different operating conditions.