Abstract
In modern electric power systems, accurately identifying short-circuit conditions is essential for reliable network monitoring and protection. This study proposes a hybrid machine learning framework for multi-class fault mode classification in three-phase electric networks. The experimental dataset included 15 operating modes, including normal mode, single-phase mode, phase-to-phase mode, three-phase mode, and ground fault mode. Input variables included phase current measurements and positive, negative, and zero-sequence current components. The preprocessing pipeline included logarithmic transformation, standardization, and correlation-based feature selection. The processed data was split into training and independent test subsets in an 80/20 ratio. Next, k-means clustering with 15 clusters was used to generate additional structural features, while random forest and extreme gradient boosting (XGBoost) algorithms created probabilistic class predictions, which were integrated by a logistic regression metaclassifier. According to the adopted testing protocol, the proposed hybrid model achieved an accuracy of 96.98%, a macro-precision of 0.973, a macro-recall of 0.969, and a macro-F1-score of 0.968, outperforming the baseline random forest model, which achieved an accuracy of 93.91% and a macro-F1-score of 0.937. These results demonstrate the potential of the proposed multi-level framework for accurate multi-class fault diagnosis in three-phase electrical networks.