MxGPS: A Multiplex Graph Transformer for Power Grid Foundation Models

Introduction

Power grids are a natural fit for graph learning, but they also expose a hard generalization problem: a model trained on one set of grid topologies may face different buses, connections, and operating structures at deployment time. The arXiv paper “MxGPS: Multiplex Graph Transformers for a Power Grid Foundation Model” frames this failure mode as topology overfitting. In the authors’ terms, task-specific training signals can lead a graph neural network to encode relational patterns tied to the training grids instead of the underlying physics.

To address this, the paper proposes MxGPS, or Multiplex GPS, a graph-transformer architecture for power-grid foundation models. It runs multiple task-specialized GPS branches over a shared node encoder and trains the system with both Static State Estimation (SSE) and AC Power Flow (PF) objectives.

Key points

• Low in-distribution error can be misleading: The paper reports that some models with substantially lower in-distribution PF error degrade sharply under topology shift, with degradation ranging from 190% to 1400%.
• Topology overfitting is the central diagnosis: Rather than blaming only insufficient capacity, the authors argue that single-task fine-tuning can over-specialize to the connectivity patterns seen during training.
• Multi-task learning acts as a constraint: By jointly optimizing SSE and PF, MxGPS forces the shared encoder to satisfy complementary gradient signals, which may reduce dependence on topology-specific shortcuts.
• Multiplex architecture: The model uses K task-specialized GPS branches on top of a shared encoder, and the paper evaluates a cross-branch attention module through ablations.
• Parameter efficiency: MxGPS is reported to use 1.6 million parameters, about 12 times fewer than the GridFM reference baseline.

Results and impact

The authors evaluate MxGPS with a three-fold sliding-window cross-validation setup spanning four unseen topologies: 14-, 24-, 162-, and 300-bus systems. In these zero-shot power-flow settings, MxGPS reportedly achieves a 0% boundary violation rate across all four topologies. The paper also states that MxGPS degrades by only 39% under topology shift, contrasting it with models that perform better in distribution but fail more severely when the grid structure changes.

The broader significance is that evaluation for grid AI models cannot stop at accuracy on familiar networks. Real power systems change because of maintenance, outages, expansion, and operational decisions. A model that memorizes structural regularities from training grids may look strong in benchmarks but remain fragile in practice.

MxGPS points toward a different design principle: foundation models for power systems should be trained not only for task accuracy but also for topology-agnostic generalization. The work is still an arXiv preprint, and its claims should be weighed against the full experimental setup, data construction, and baseline choices. Even so, it highlights an important lesson for AI in scientific infrastructure: smaller, physics-aware, multi-task models may generalize more reliably than larger single-task models optimized only for in-distribution metrics.

Source: arXiv