Topology-Aware Reinforcement Learning for Tertiary Voltage Control

Transmission systems have experienced an increase in the occurrence frequency and intensity of high voltage events over the past few years. Since traditional approaches to optimal power flow do not scale well to real-life systems, it has become urgent to develop new methods to help operators improve tertiary voltage control. In this paper, we propose to train a graph neural network to choose voltage setpoints by interacting with a power grid simulator using reinforcement learning techniques. Moreover, we introduce the hyper heterogeneous multi graph formalism to account for topology variations of real-life systems (assets disconnection, bus-splitting, etc.). Our approach is validated on an artificial case study based on the case60nordic power grid.