Digital Twin Integrated Safe Deep Reinforcement Learning for Resilience Enhancement of Unbalanced Distribution Grids

The growing integration of distributed energy resources (DERs) into distribution grids and increasing extreme events bring security challenges into grids, while accurate grid model is difficult to be obtained due to the widespread deployment of DERs and unknown extreme events. This paper proposes a new method to strengthen the resilience of unbalanced distribution networks through the integration of a novel power system digital twin (DT) and deep reinforcement learning (DRL) based multi-step critical load restoration method. First, a combined data driven and model based identification method is proposed to develop power system DT, which achieves the identification of network topology and line characteristics of distribution grids even under unbalanced conditions. Moreover, a safe DRL based multi-step critical load restoration scheme with dynamic microgrids (MGs) formation capability is established to improve distribution system resilience after extreme events and realize zero voltage violations. Finally, a DT integrated DRL based resilience enhancement approach is proposed to achieve real-time model updating, minimal load loss and guaranteed grid safety under unbalanced and uncertain grid conditions.