The proliferation of hybrid distribution networks combining overhead lines and power cables introduces pronounced non-uniformity in line parameters, leading to high spatiotemporal complexity in transient fault behavior. To address the challenges of weak, noisy, and highly nonlinear high-impedance ground fault signals, this paper proposes a robust multi-modal deep fusion framework for accurate ground fault identification and localization. The framework integrates a Temporal Convolutional Network (TCN), Bidirectional Gated Recurrent Unit (BiGRU), and attention mechanisms to capture multi-scale temporal dependencies and enhance key transient features from multi-source electrical signals. To overcome the feature misalignment between heterogeneous modalities, a Quantized Residual Transformer (QRTransformer) is constructed. It introduces a quantization-based regularization mechanism to suppress noise propagation while facilitating deep semantic interaction among voltage, current, and decomposed intrinsic mode functions (IMFs). Furthermore, a probabilistic knowledge optimization (PKO) strategy combined with imprecise probability theory enables uncertainty-aware classification and significantly improves robustness under small-sample and noise-disturbed conditions. Extensive simulations and closed-loop hardware-in-the-loop (HIL) experiments, utilizing an OPAL-RT real-time simulator and an ARM-based embedded edge terminal, demonstrate that the proposed framework achieves a diagnostic accuracy of 96.8% and an average intra-section fault distance estimation error of 0.18 km, conditioned on correct section identification, outperforming conventional approaches in both precision and generalization. The HIL validation confirms that the modular architecture supports low-latency real-time deployment in smart substations or edge computing terminals, offering a scalable paradigm for intelligent diagnosis in complex hybrid distribution networks.
Yunwang et al. (Tue,) studied this question.