A reinforcement learning approach to parameter selection for distributed optimal power flow

With the increasing penetration of distributed energy resources, optimization algorithms have attracted significant attention for power systems applications due to their potential superior scalability, privacy, and robustness a single point-of-failure. The Alternating Direction Method Multipliers (ADMM) is popular algorithm; however, its convergence performance highly dependent on selection penalty parameters, which are usually chosen heuristically. In this work, we use reinforcement learning (RL) develop an adaptive parameter policy alternating current optimal flow (ACOPF) problem solved via ADMM with goal minimizing number iterations until convergence. We train our RL using deep Q-learning show that can result in significantly accelerated (up 59% reduction compared existing, curvature-informed methods). Furthermore, demonstrates promise generalizability, performing well under unseen loading schemes as losses lines generators 50% iterations). This work thus provides proof-of-concept applications. • targets task ACOPF problems ADMM. propose adaptively adjust parameters. Parameters produced by accelerates Trained generalizes varying loads network outage.