Hybrid AI-driven intelligent fault diagnosis and localization in modern power systems

Abstract

This paper presents a hybrid intelligent framework for fault diagnosis and localization in modern power distribution systems, addressing challenges such as noisy measurements, high-impedance faults (HIF), and uncertain operating conditions. The proposed approach integrates deep neural networks (DNN) for nonlinear feature extraction, support vector machines (SVM) for robust classification, and a fuzzy inference system for uncertainty-aware decision fusion, combining the strengths of deep learning, machine learning, and soft computing. A comprehensive dataset of over 12,000 fault instances is generated using IEEE 33-bus and 69-bus systems, covering multiple fault types (LG, LL, LLG, LLL), fault resistances (0.1-200 Ω), varying load conditions, and noise levels from 30 dB to -5 dB SNR. Wavelet-based denoising and hybrid feature extraction (time–frequency and statistical features) are employed to capture transient characteristics. The DNN generates discriminative feature embeddings, which are classified using an RBF-kernel SVM and further refined through fuzzy logic with Gaussian membership functions. Fault localization is performed using impedance-based estimation enhanced by learned correction. Results show that the proposed model achieves 98.5% classification accuracy, outperforming DNN (93.2%), SVM (90.4%), random forest (91.1%), and k-NN (88.6%). The model demonstrates strong noise robustness, with only -6% accuracy degradation at -5 dB SNR. It achieves fault localization error of 0.2-0.7 km and HIF detection with F1-score of 0.91. With inference latency of 45 ms (reduced to 28 ms after optimization), the system is suitable for real-time deployment, providing a scalable and reliable solution for smart grid fault monitoring.