Physics-Aware Distillation of a Grid Foundation Model for Real-Time Edge Inference

Key points

• Their size, however, targets data-centre inference, whereas low-inertia operation increasingly asks for inference at the edge, next to the measurement, at batch size one and within a single synchrophasor (PMU) reporting frame.1
• We distil a 20.1M-parameter, 12-layer GridFM teacher into a 0.62M-parameter, 6-layer student, a $32.5\times$ parameter reduction at half the message-passing depth, using an objective that combines the masked power-flow loss with output, physics-residual, latent-state and attention transfer.1
• The physics-residual term, which matches the teacher's power-residual profile, has not previously been applied to typed power-grid graphs.1
• The physics-residual term, in the layer-aggregated form used here, is consistently harmful, which we attribute to the spatial averaging it applies to a bus-level residual field.1

Sources (1)

• [1]Physics-Aware Distillation of a Grid Foundation Model for Real-Time Edge Inference
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 15, 01:15 PM

  Grid foundation models (GridFMs) are graph transformers pretrained on a masked power-flow objective, and evaluate power flow orders of magnitude faster than an iterative solver.

  Their size, however, targets data-centre inference, whereas low-inertia operation increasingly asks for inference at the edge, next to the measurement, at batch size one and within a single synchrophasor (PMU) reporting frame.

Extractive summary: sentences quoted from the sources.