Optimized Machine Learning Approach for Accurate Detection of Sub-Synchronous Oscillations in Wind Turbines

Sub-synchronous oscillations (SSO) create a serious risk to equipment safety and grid stability, and their occurrence has increased with the growing integration of renewable energy sources. Therefore, the implementation of efficient mitigation strategies depends on accurate, timely detection of SSO. To forecast and evaluate the dynamic stability of wind turbine systems with respect to SSO, this study presents a machine learning (ML) approach. The IEEE First Benchmark Model (FBM) was used to generate dynamic data, consisting of 10,648 scenarios generated by varying wind speed (WS), the number of turbines in service (WT), and series compensation (SC) levels. The model also uses input features such as active and reactive power, voltage, current, and torque to classify the system's SSO state into four categories: No SSO, Non-Severe SSO, Non-Critical SSO, and Severe SSO. With a peak accuracy of 91.38 %, the Decision Tree (DT) model outperformed the other models in the test. The proposed approach offers wind farm operators a quick and reliable tool for stability assessment and compensation planning. Additionally, optimizing the traintest data split using metaheuristic algorithms (WOA, GA, and PSO) improved the predictive performance of the ML models. After optimization, the DT model's accuracy exceeded 92%, highlighting the importance of strategic data partitioning for capturing complex SSO patterns.