Dynamic safety-guaranteed control for direct current microgrids via safe reinforcement learning

To address voltage instability and safety risks under large disturbances in grid-forming direct current microgrids with constant power loads, this paper presents a dynamic safety-guaranteed nonsmooth control strategy with soft actor-critic–based online parameter adaptation and control barrier function-based quadratic programming–based safety filtering. A baseline nonsmooth controller is first designed to guarantee robust voltage regulation as the nominal control law. To explicitly enforce transient-current safety, a control barrier function-based quadratic programming layer is introduced to modify the nominal nonsmooth control input only when necessary, thereby constraining the inductor current within the prescribed safety boundary during fast load variations. On this basis, the soft actor-critic algorithm is employed to adapt key nonsmooth controller parameters online rather than directly generating the converter duty cycle. Simulation and experimental results demonstrate that the proposed controller maintains stable and safe operation under constant power load–induced disturbances and parameter uncertainties.