A weighted ensemble learning framework for data-driven fault detection and classification in active distribution networks

Abstract

The increasing integration of distributed energy resources (DERs) complicates fault detection in active distribution networks, rendering traditional model-based approaches inadequate for handling complex, time-varying power system characteristics. This paper proposes a novel weighted ensemble learning framework integrating multiple heterogeneous classifiers—Random Forest, XGBoost, Support Vector Machine, Multi-Layer Perceptron, and K-Nearest Neighbors—for accurate and robust fault classification. The ensemble weights are optimized through 5-fold cross-validation accuracy to effectively leverage the complementary strengths of diverse decision boundaries. A dataset of 6,000 samples was generated via OpenDSS time-domain simulations on the IEEE 33-bus test feeder, incorporating 30% DER penetration and measurement noise across five fault categories with 12 electrical features. Extensive experiments demonstrate that the proposed ensemble achieves a superior accuracy and F1-score of 98.83%, significantly outperforming individual classifiers and deep learning baselines, including 1D-CNN (94.75%) and LSTM (97.42%). Ablation studies and robustness analyses confirm its practical applicability. Finally, detailed confusion analysis reveals that overlapping zero-sequence and negative-sequence signatures constitute the main misclassification source, motivating future work on high-frequency transient feature extraction.