Diseño de protecciones para sistemas aislados de tierra usando algoritmo de detección de fallas con IA

Abstract

CONDITIONING OF PROJECT PUBLICATION. intelligent scheme is proposed that uses machine learning to recognize ground faults in milliseconds, which often go unnoticed in isolated neutral networks. The aim is to reduce clearance time and increase operational continuity, based on the hypothesis that a lightweight classifier can outperform fixed logic without adding expensive hardware. To test this hypothesis, an IEEE 14-bus system was modeled in MATLAB-Simulink. More than 30,000 scenarios were simulated with load variations and random short circuits, and the primary currents were scaled to secondary values using typical current transformer ratios. The RMS magnitudes, peaks, and sequence components were extracted and used to train a random forest of 300 trees; the model was then validated with unpublished files. Single-phase, two-phase, and three-phase faults were classified with a confidence index greater than 0.96 and a latency of 16 to 25 ms, while false trips remained below 2%. It is concluded that the proposal offers fast, adaptable, and cost-effective detection for critical networks operating without a connected neutral. Keywords: Isolated system, Transients, Random Forest, Simulink, Transformers