Modelling of Switching Operation Process and Real-Time Violation Detection Method Based on Dynamic Graph Transformer

Abstract

Substation switching operation is a high-risk manual procedure where any misordering or omission may cause serious safety accidents. Existing rule engines cannot cover the diverse violation patterns induced by topology switching, and sequence-based deep models ignore electrical topology constraints and the ticket-level operation task. This paper proposes a real-time detection method that models the switching-operation process as a dynamic heterogeneous graph and encodes it with a spatiotemporal dual-channel dynamic graph Transformer. Each step triggers an incremental hard-update of graph topology, producing snapshots enriched with device state, electrical polarity, voltage level and topological position; the spatial graph-attention channel and the temporal causal Transformer channel are jointly conditioned on the operation-task label. A contrastive strategy mixing controlled violation injection with real held-out violations alleviates label scarcity; a task-conditioned multiprototype scorer captures the multimodal distribution of normal operations; a three-level risk-graded thresholding scheme aligns alarms with safety-consequence severity; and a ticket-level task-consistency check catches violations where all steps are locally correct but the ticket does not match the declared task. Experiments on 6842 tickets from 12 real substations achieve F1 = 87.4% (+2.0% over the best 2025 SOTA baseline), 0.6-step detection latency, 18.3-ms single-step inference, 91.2% top-3 step localisation and only 3.6% F1 drop on three unseen substations.