Neural Networks for AC Optimal Power Flow: Improving Worst-Case Guarantees during Training

Abstract

The AC Optimal Power Flow (AC-OPF) problem is central to power system operation but challenging to solve efficiently due to its nonconvex and nonlinear nature. Neural networks (NNs) offer fast surrogates, yet their black-box behavior raises concerns about constraint violations that can compromise safety. We propose a verification-informed NN framework that incorporates worst-case constraint violations directly into training, producing models that are both accurate and provably safer. Through post-hoc verification, we achieve substantial reductions in worst-case violations and, for the first time, verify all operational constraints of large-scale AC-OPF proxies. Practical feasibility is further enhanced via restoration and warm-start strategies for infeasible operating points. Experiments on systems ranging from 57 to 793 buses demonstrate scalability, speed, and reliability, bridging the gap between ML acceleration and safe, real-time deployment of AC-OPF solutions - and paving the way toward data-driven optimal control.