Lightweight AI-Based Fault Locator for Renewable Based Power Systems

This study presents two AI-based models for detecting, classifying, and locating short-circuit faults in power grids that include renewable energy sources: a standard Multi- Layer Perceptron (MLP) and a more advanced Kolmogorov- Arnold Network (KAN). Both perform well in detecting faults and classifying their types. However, KAN consistently outperformed MLP in locating faults, especially under challenging conditions like non-grounded faults and high fault resistance. The study also shows that using a large and diverse set of engineered features, such as time-domain metrics, wavelet coefficients, and sequence components, enhances model performance. Overall, KAN proves to be a more capable and flexible model, offering reliable results across all tasks using only local data. This makes it a promising tool for modern grid protection in systems with high renewable integration.