Two-timescale online coordinated schedule of active distribution network considering dynamic network reconfiguration via bi-level safe deep reinforcement learning

Mass access of renewable energy leads to high voltage fluctuations and network loss. To coordinate different scheduling resources in different timescales while improving voltage quality and reducing network loss, a two-timescale coordinated model via a bi-level safe deep reinforcement learning (DRL) algorithm is designed. Based on pre-generated topological feasible domains, the upper layer Rainbow algorithm determines the network topology as well as the switching gears of the OLTC and SC in the slow timescale. While the lower layer continuous safeDDPG algorithm schedules the distributed photovoltaic inverters in the fast timescale based on the chosen actions of the upper layer. The obtained power flow results will be used to continuously update the upper and lower network parameters, thus realizing the coordination of different equipment in two timescales. Numerical results show that the proposed bi-level Rainbow-safeDDPG algorithm can effectively solve the voltage and network loss issues and stabilize the nodal voltage within the safe range. Both voltage violation and network loss cost are significantly reduced due to the adoption of network reconfiguration method. Compared to traditional DRL algorithms, the training efficiency is improved and enables secure exploration and training. Finally, the large-scale 116-nodes testing system verifies the scalability of the algorithm.