Complete ML-Based Methodology for Fault Classification in Onshore Wind Farm Collector Networks

The proliferation of onshore wind farms, characterized by complex collector networks and inverter-based resources (IBRs), introduces significant challenges for conventional fault diagnosis systems. Given these issues, this paper presents a robust machine learning (ML)-based methodology for accurate fault classification tailored for wind farm collector systems. The proposed approach relies on features derived from incremental voltage signals to enhance robustness against noise and complex control behaviors inherent to IBRs. The methodology encompasses a multistage process, starting with the evaluation of 24 different classifiers, followed by a systematic feature selection process to improve noise immunity and comprehensive hyperparameter optimization using both Grid Search and a custom-designed Genetic Algorithm. The finalized models are rigorously validated against a wide range of parameters, including fault resistance, inception angle, and location, as well as the varying operating power factor of the IBR, the penetration level of the generation, and the signal-to-noise ratios. The results of 55800 simulated fault scenarios demonstrate the superiority of the proposed framework. The optimized Extra Trees model, in particular, achieved exceptional accuracy, exceeding 97% accuracy even at a challenging 40 dB noise level, demonstrating its effectiveness for real-world applications. Beyond providing a highly accurate and robust solution for fault classification in collector systems with IBR, this work delivers a complete benchmark methodology that can serve as a reference for future development and evaluation of ML-based diagnostic tools in power systems.