Physics-grounded sequence-domain graph learning for adaptive fault location in IBR-dominated offshore wind farm transmission networks

Fault location in offshore wind-farm networks dominated by inverter-based resources (IBRs) is complicated by converter-limited fault currents, cable charging currents and high fault resistance, which violate the assumptions of classical impedance-based estimators. This paper proposes PG-SDGL, a physics-grounded sequence-domain graph-learning framework that combines a graph-attention sequence-domain graph neural network (SeqGNN), a three-sequence physics-informed neural network (3Seq-PINN) for fault-resistance decoupling, sequence-residual split-conformal prediction intervals, cable charging-current compensation, an analytical IBR fault-current decomposition, a harmonic high-impedance-fault (HIF) gate with Dempster–Shafer fusion of two intelligent electronic devices and a protection-consistency index (PCI). The framework was evaluated in simulation on a 69 kV network with a 60 km overhead interconnection line and three 5–9 km submarine cable feeders, for four fault types and fault resistances of 2–150 Ω. Feeder-identification accuracy was 99.4–99.6%. Over 2–40 Ω, the average main-line mean relative error (MRE) was 0.195%, compared with 0.631% in the published random-forest benchmark on the same line model. The average fault-resistance mean absolute error decreased from 4.67 Ω (random-forest benchmark) to 2.07 Ω (single-sequence PINN) and 1.98 Ω (3Seq-PINN). The sequence-residual intervals reached 95.0–96.9% empirical coverage with a 33.8% smaller average normalized width than plain split-conformal intervals. At 150 Ω, HIF detection accuracy was 97.4%, location MRE was 1.48% and PCI was 0.924.