Improved Deep Reinforcement Learning with Logistic Map for Microgrid Energy Management

Providing the optimal solution while reducing computational time in solving the energy management task is more challenging for developing an Energy Management System (EMS) of a Microgrid (MG), especially when considering cost minimization and protection against MG constraint violations. A state-of-the-art approach, a deep reinforcement learning is mostly implemented as a smart agent for optimally controlling a battery installed in the MG, especially Deep Deterministic Policy Gradient (DDPG). This helps address uncertainties in electricity generation from distributed energy resources and electricity consumption by related loads. However, the structure of the DDPG agent, which uses neural networks with many hyperparameters, often leads to non-convergence and high computational burdens during the training process. To address these issues, DDPG is improved by assisting a Logistic Map (LM) which is proposed in this work. The LM is employed to find better initial weights for the agent and to enhance exploration during the training process. The experimental results demonstrate that LM-DDPG outperforms conventional DDPG by providing faster convergence, offering a better solution for battery control, and reducing lower daily total cost by 13.38%.