Reliability-driven distribution power network dynamic reconfiguration in presence of distributed generation by the deep reinforcement learning method

Abstract

The reliability of the distribution network increasingly common by high penetration of distributed generations. To address the reliability, this paper propose a new approach based on the dynamic reconfiguration to improve the reliability of the distribution network in presence of distributed generation by Reinforcement Learning algorithm. In this research, the deep reinforcement learning approach find the optimal switches to change the power flow when the network reliability is reduced. The objective function as a reward in reinforcement learning algorithm included active power loss, voltage deviation, and reliability. Furthermore, the failure-rate reduction strategy was computed as a reliability indices of the load points in the distribution network. To calculate the probabilistic load flow, the 2m Point Estimate Method (PEM) has been proposed. The proposed approach is evaluated using the IEEE 33-bus and the IEEE118-bus. The suggested structure has been implemented by deep Q-learning method. The Guaranteed Convergence Particle Swarm Optimization (GCPSO) and Honeybee-Graph (HG) algorithms has been implemented to comparison with presented approach. The results illustrate the usefulness of the methodology. By comparison, near to 5, 10 and 4, 14 percentage improvement by DQL method than the HG, GCPSO methods in total reactive and active power loss, respectively, in the IEEE 118 bus test system (for all scenarios). Furthermore the computing time for presented method is less than the GCPSO and HG.