Physics-Informed Neural Networks for AC Optimal Power Flow

This paper introduces, for the first time to our knowledge, physics-informed neural networks accurately estimate AC-OPF result and delivers rigorous guarantees about their performance. Power system operators, along with several other actors, are increasingly using Optimal Flow (OPF) algorithms a wide number of applications, including planning real-time operations. However, in its original form, AC problem is often challenging solve as it non-linear non-convex. Besides large approximations relaxations, recent efforts have also been focusing on Machine Learning approaches, especially networks. So far, however, these approaches only partially considered physical models available during training. And, more importantly, they offered no potential constraint violations output. Our approach (i) introduces power flow equations inside network training (ii) integrates methods that rigorously determine reduce worst-case across entire input domain, while maintaining optimality prediction. We demonstrate how achieve higher accuracy lower than standard networks, show we can further all