Q-Learning Routing Protocol Applied to Energy Internet for Power Distribution

Several routing methods for power distribution have been presented in the literature. However, the majority either have a high computation time and complexity or do not support congestion management and dynamic networks. As a solution, this paper proposes a novel power routing protocol that allows congestion management, supports dynamic networks, and has low computation time and communication overhead. The power routing problem is treated as a distributed control and learning problem based on the multi-agent reinforcement learning (MARL) approach. We aim to find the best routing policy for minimizing power losses by supporting all power constraints. The proposed protocol is implemented in each energy router (ER) and requires only local information to choose the next router to transmit a power packet (PP) for a specific destination. Simulation results show that the Q-learning-based power routing protocol has low communication overhead (only local information is required), low computational time, and supports dynamic change in the power network. Finally, it outperforms Dijkstra’s algorithm in terms of computation time, communication overhead, and application in a dynamic environment.