Ferroresonance overvoltage identification in distribution networks based on Markov transition field and GNN–Dual-ViT: A Markov–graph co-design framework

Ferroresonance is a typical abnormal operating condition in distribution networks. The overvoltages it produces can damage potential transformers (PTs) and, in extreme cases, cause switchgear explosions and electrical fires. The wide spectrum and strong nonlinearity of ferroresonance signals render conventional time–frequency methods inadequate. This paper proposes a Markov–graph co-design framework that tightly couples a theoretically grounded encoding stage with a structure-aware classifier. First , the first-order Markov approximation for quantised ferroresonance sequences is assessed using conditional mutual information and a conditional-independence test, with partial autocorrelation and Ljung–Box residual diagnostics used as supporting evidence. The combined results provide an empirical basis for Markov Transition Field (MTF) encoding without imposing an ordinal interpretation on the categorical states. Second , MTF converts one-dimensional voltage signals into weighted directed graphs whose edge weights encode state-transition probabilities. Third , a GNN–Dual-ViT architecture with progressive multimodal fusion exploits the resulting graph topology. The framework is evaluated against seven representative baselines reimplemented on a common ATP–EMTP dataset under identical protocols with multiple random seeds, and is further stress-tested under low SNR and load fluctuation. A cross-voltage study distinguishes same-mechanism transfer to a 35 kV PT system from adaptation to a physically different 110 kV CCVT regime. The results indicate that transfer is feasible with lightweight domain adaptation and limited labelled target data, while regime-specific calibration remains necessary for deployment. A branchwise qualitative analysis, a failure-case taxonomy, and a maximum-mean-discrepancy audit of the sim-to-field domain gap characterise both the mechanism and the boundary of applicability of the proposed framework.