Reproduce “Scalable Self-Supervised Learning for Multiphase AC-OPF in Distribution Systems with Topology Reconfiguration” on OpenResearch

About this paper

Hoang T. Nguyen, Shaohui Liu, Reetam Sen Biswas, Varsha Pendyala, Nurali Virani, Deepjyoti Deka, Priya L. Donti

A self-supervised surrogate combines differentiable multiphase power-flow evaluation, topology updates via Sherman–Morrison–Woodbury, and sequential linearized feasibility seeking to approximate AC-OPF across reconfigurable distribution feeders. On IEEE systems ranging from 13 to 8,500 nodes, the method reduces constraint violations to as low as 3.2×10^-8, keeps optimality gaps below 1.5%, and achieves inference speedups of up to 1,121× over IPOPT.

Key insights

• Embedding physical power-flow computation is more reliable than directly predicting voltages: the direct penalty baseline failed to converge across the tested feeders, whereas the proposed approach maintained low violations.
• Feasibility repair complements rather than replaces the learned predictor, reducing violations by one to three orders of magnitude with only modest additional cost error.
• The combination of low-rank topology updates and truncated differentiation addresses two major scaling bottlenecks, with three-step differentiation reported as 2.74× faster than full unrolling on the 906-bus feeder.

Summary by alphaXiv.