Graph Neural Networks for Transmission Grid Topology Control: Busbar Information Asymmetry and Heterogeneous Representations

Source

Publication And Credibility

  • Paper date: arXiv published 2025-01-13; v3 updated 2025-10-03.
  • Venue/status: arXiv preprint.
  • Credibility: TenneT/Radboud authors and practical TSO framing make it credible applied evidence, but it remains a preprint unless a venue version is verified.

Core Claim

The paper identifies busbar information asymmetry in common homogeneous graph representations and proposes a heterogeneous grid graph representation that improves GNN topology-control effectiveness and OOD topology generalization.

L2RPN / Grid2Op Notes

The experiments use Grid2Op rte_case14_realistic, planned/unplanned outages, and N-1 network variants. The key contribution is graph representation hygiene for topology-control agents, not a new RL algorithm.

Action-Time-Series / World-Model Notes

For action-conditioned world models, this is a schema warning: a graph latent that omits potential busbar connectivity can fail to preserve the variables needed to predict action consequences under topology changes.

Foundation TSFM Relevance

Agenda slotVerdictEvidenceMissing pieces
Context interface: topology and channel contextpartially closesShows that heterogeneous graph representation can matter for topology-control generalization.Needs validation on larger grids and multi-step transition objectives.
Causal structure, counterfactuals, and controladjacentImitation/control evaluation depends on topology actions.No learned dynamics model or candidate-action rollout.
Benchmark hygienepartially closesExplicitly tests OOD network configurations.Remains focused on rte_case14_realistic.