A graph policy network approach for Volt-Var Control in power distribution systems

Volt-var control (VVC) is the problem of operating power distribution systems within healthy regimes by controlling actuators in systems. Existing works have mostly adopted conventional routine representing (a graph with tree topology) as vectors to train deep reinforcement learning (RL) policies. We propose a framework that combines RL neural networks and study benefits limitations graph-based policy VVC setting. Our results show policies converge same rewards asymptotically, however at slower rate when compared vector representation counterpart. conduct further analysis on impact both observations actions: On observation end, we examine robustness two typical data acquisition errors systems, namely sensor communication failure measurement misalignment. Furthermore, erroneous topological information representation. reveal are significantly more robust than dense network representations. action various impacts system, thus using induced physical topology may not be optimal choice. In case demonstrate choice readout function architecture augmentation can improve training performance robustness.