Scaling Dual Conic Proxies for AC Optimal Power Flow

Optimization proxies, machine learning models that map parameters of optimization problem instances to solutions, have the potential to significantly increase the efficiency of real-time decision-making, even unlocking conventionally intractable applications, by relegating online computation to offline training. However, unlike trusted optimization algorithms, optimization proxies do not provide optimality certificate of the solutions they predict. To address this gap, we proposed the dual conic proxies (DCP), a framework of optimization proxies that predict valid optimality certificates. We have applied this methodology to AC optimal power flow (OPF), a nonconvex optimization problem central to power system operations. In this workshop paper, we present new results on improving the certificate quality and scalability of such DCP models.