Wavelet-based fault detection and classification in MV grids using neural models: case study of Moroccan distribution system

Abstract The increasing complexity of modern power distribution systems, driven by renewable integration, evolving load patterns, and aging infrastructure, has accentuated the need for advanced fault detection and classification mechanisms, particularly in emerging medium-voltage (MV) networks such as Moroccan distribution grid. However, traditional protection schemes, often based on centralized logic and fixed thresholds, tend to underperform in complex or high-impedance fault conditions. Furthermore, global signal features such as RMS or frequency components are insufficient to capture the localized and phase-dependent behavior of faults. These limitations have prompted a growing interest in intelligent, data-driven approaches combining signal processing and machine learning to achieve high-resolution fault diagnosis and improved system reliability. The present study proposes and evaluates an intelligent fault classification framework tailored to MV distribution networks. It explores the comparative performance of three neural architectures and supervised learning classifiers, Multilayer Perceptron (MLP), Support Vector Machine (SVM) and Radial Basis Function Neural Network (RBFNN), applied to both globally extracted features and phase-localized wavelet descriptors. In addition to the baseline classification framework, a novel phase-based analysis method is introduced to enhance diagnostic performance. This method processes each phase and neutral current independently, based on the hypothesis that fault signatures emerge more distinctly when analysed separately. The experimental results demonstrate that neural models can reliably identify fault types across both approaches and underscores the potential of wavelet-enhanced AI models for smart and localized protection in emerging power distribution systems and prove its readiness for real-world deployment in SCADA-integrated protection systems.