Architectural Vulnerability Mapping and Analysis in Hybrid AC/DC Distribution Networks

Abstract

The growing integration of distributed renewable resources is pushing medium-voltage distribution grids toward hybrid AC/DC configurations, whose topology-dependent vulnerabilities, including harmonic resonance, sub-synchronous & converter-control interactions (SSCI/CCI), AC/DC interface faults, and protection miscoordination, remain poorly addressed. This paper develops a vulnerability mapping framework for hybrid MVAC/MVDC systems that combines multilayer network topology modelling, multi-source harmonic-domain analysis using frequency coupling matrices, and Backward-Forward Sweep power flow. The relationship between topology, resonant-mode count, and SSCI phase-margin coupling is derived analytically from the admittance matrix. The framework is applied to 12 configurations across four topologies, using phasor-domain simulation with electromagnetic transient (EMT) verification of fault and converter cases. The results show that, for an XLPE 95 mm2 DC cable, the dominant resonance falls below 200 Hz for feeders longer than about 12.6 km; Monte Carlo analysis places this crossing at 10.0–14.2 km (5th–95th percentile). At 60% renewable penetration, four configurations (C5, C8, C9, C12) exceed the unity resonance-vulnerability threshold, and the meshed DC-side grid-forming configuration (C8) violates the 30° phase-margin limit. Because the penetration at which this limit is lost ranges from 20% to 85%, grid-forming deployment should rely on configuration-specific impedance and phase-margin assessment rather than a universal threshold. Overcurrent relays misoperate in 31–61% of DC pole-to-ground faults, reflecting a mismatch between magnitude-based protection and DC fault signatures under high-impedance grounding. A long short-term memory (LSTM) classifier trained on 468 simulated scenarios achieves 94.4% accuracy on 120 held-out scenarios under noise-free conditions, but its RVI, SVI, and FPSS accuracy fall to 42.5–50.8% when the meshed topology class is withheld, so each topology class must be represented in the training data.