Abstract: The randomness and intermittency of distributed energy have brought great challenges to the energy management of direct current (DC) microgrids. In order to solve this challenge, a DC microgrid energy management strategy based on reward guidance deep deterministic policy gradient (RG-DDPG) is proposed in this paper. This strategy describes the optimal operation of the DC microgrid as a Markov decision process and uses the continuous interaction between the agent and the DC microgrid environment to adaptively learn energy management decisions, thus realizing the optimal management of the DC microgrid energy. In the strategy training process, the priority experience replay mechanism based on temporal difference error (TD-error) is used to reduce the randomness and blindness of RG-DDPG’s learning and exploration in the DC microgrid operating environment and improve the convergence speed of the energy optimization and management strategy proposed in this paper. At the same time, during the training rounds, the size of the accumulated rewards is used to construct an excellent round set of DC microgrid energy management, strengthen the connection between RG-DDPG agents in the training rounds, and maximize the use of the training value of the excellent round. The simulation results show that the proposed strategy can reasonably distribute energy in the DC microgrid. Compared with the energy management strategy based on deep Q learning (DQN) and particle swarm optimization (PSO), the proposed strategy can reduce the daily average operation cost of DC microgrids by 11.16% and 7.10%, respectively.

Key words: DC microgrid, energy management, RG-DDPG, priority experience replay, excellent round set