Reinforcement Learning using Physics Inspired Graph Convolutional Neural Networks

In this work, we propose a physics inspired Graph Convolutional Neural Network (GCN)-Reinforcement Learning (RL) architecture to train online controllers policies for the optimal selection of Distributed Energy Resources (DER) set-points. While the use of GCN is compatible with any DRL scheme, we test it in combination with the popular proximal policy optimization (PPO) algorithm and, as application, we consider the selection of set-points for Volt/Var and Volt/Watt control logic of smart inverters as the case study for DER control. We are able to show numerically that the GCN scheme is more effective than various benchmarks in regulating voltage and mitigating undesirable voltage dynamics generated by cyber-attacks. In addition to exploring the performance of GCN for a given network, we investigate the case of grids that are dynamically changing due to topology or line parameters variations. We test the robustness of GCN-RL policies against small perturbations and evaluate the scheme so called "transfer learning" capabilities.