Constrained Reinforcement Learning-Based Framework for DC Microgrid: Current Sharing and Voltage Regulation

In this paper, a constrained reinforcement learning-based framework is developed for islanded DC microgrids to deal with the issues of voltage regulation and current sharing. First, a new local Q function including the current error and voltage error is designed on the basis of a hyperbolic function. The solution of the HJB equation associated with the local Q function yields an optimal policy. Then, a novel saturation-input policy iterative (SIPI) method is designed to approximate this policy. Its convergence is also analyzed in detail. To implement this SIPI method, an actor–critic control structure is developed, in which the actor network employs an asymmetric bounded function as its activation function to handle asymmetric input saturation. Finally, the effectiveness of the SIPI method is verified through numerical simulations.