Dimension Reduction Techniques for Machine Learning-Based AC Microgrid Fault Diagnosis: A Systematic Review

The integration of inverter-interfaced sources into microgrids has led to new protection challenges due to fluctuating fault currents, bidirectional power flow, and low inertia. Traditional protection methods are inadequate for these evolving systems, necessitating the development of novel, fast-acting, and cost-effective strategies that are independent of fault current levels. Machine learning-based protection schemes offer potential for improved performance over traditional methods, but their success is highly dependent on the quality and nature of the input data. Raw fault signal data are often noisy and high-dimensional. Processing it without feature extraction or selection can result in poor performance, longer processing times, higher storage needs, overfitting, and reduced generalizability. The dimensionality reduction step, which encompasses feature extraction and/or feature selection algorithms, is crucial for machine learning-based protection schemes. Many such methods have been proposed in the last decade. This research aims to investigate available techniques for reducing the dimensionality of data used for fault diagnosis in AC microgrids. The research will review and compare different dimensionality reduction methods, highlighting their strengths and weaknesses. It will also explore the most used features and their effectiveness in different scenarios. This article highlights the key gaps in AC microgrid fault diagnosis, which include scalability, real-time performance, and generalization, among many others. It emphasizes that future research should focus on unexplored dimension reduction techniques, hybrid approaches, and improving deep learning interpretability for robust and accurate solutions.