Real-time operation of distribution network: A deep reinforcement learning-based reconfiguration approach

With the growing penetration of renewable energy sources and remotely controllable switches, reconfiguration processes have increasingly played a critical role in the real-time operation of distribution networks. A dynamic network reconfiguration strategy is developed in this study with the objective of minimizing both the operation cost and the amount of load shedding. A three-stage deep reinforcement learning-based optimization method is proposed to determine the optimal reconfiguration and set-points of distributed generators in real-time operation. In stage 1, day-ahead scheduling is carried out for unit commitment and economic dispatch. In stage 2, the optimal configuration is determined to consider different events in the distribution system. These events are usually the disconnection of various branches or buses in the system to isolate the faults in different locations. In stage 3, the real-time set-points of distributed generators are determined considering uncertainties. Using deep neural networks as function approximators, the proposed method is able to find out the optimal configuration and real-time set-points immediately. The fast response of the proposed method enhances the stability and service reliability of distribution networks. The performance of various reinforcement learning algorithms is also analyzed to determine the best method for the proposed strategy. A microgrid and IEEE 33-bus networks are used to validate the performance of the proposed method.

Automated summary

This study develops a dynamic network reconfiguration strategy to minimize operation cost and load shedding by using deep reinforcement learning for determining the optimal reconfiguration and set-points of distributed generators in real-time operation. The proposed method improves the stability and service reliability of distribution networks by quickly finding the optimal configuration and real-time set-points.