UNION: Topology-Conditioned AC-OPF under Structural Grid Shifts

Learned AC optimal power flow (AC-OPF) is commonly evaluated on load perturbations of a fixed network, leaving reliability under structural grid shifts unclear. We present UNION, a graph-to-control-to-physics framework that shares an encoder across systems, predicts bounded generator controls, solves the active-topology nonlinear AC equations through a sparse differentiable implicit layer, and applies deterministic restoration to residual inequality violations. One model is jointly trained on seven heterogeneous systems, including a proprietary 4,492-bus transmission grid, and evaluated on 14,000 held-out normal-operation instances, unseen $N-1$ line and generator outages, and 116 hourly snapshots with changing topology and generator availability. UNION achieves a 1.23% delivered objective gap and 99.56% strict instance feasibility in normal operation; across zero-shot contingencies, these metrics range from 0.24--1.74% and 70.07--100%, while temporal fine-tuning yields 100% coverage, a 2.51% gap, and 82.8% feasibility. Failures reflect limited voltage control, redispatch leverage, and reference-generator balancing, motivating joint reporting of full-AC feasibility, coverage, and failure stage.