A DRL-Based Parameter Self Configuration Mechanism of Nonsmooth Control for Autonomous DC Microgrids Feeding Constant Power Loads

To accommodate constant power loads (CPLs) with varying degrees of disturbances levels in dc microgrid systems, the adaptability of existing robust control strategies should be guaranteed. Regarding this issue, this article proposes a novel parameter self-configuration mechanism based on deep reinforcement learning (DRL) in an effort to achieve a convincing balance between the robustness and adaptability of a large-signal stabilizer. First, a recent developed nonsmooth composite controller with a fractional order fine-tuning robustness factor is employed to stabilize the dc microgrids. Second, the Markov decision process (MDP) is formulated according to the control objectives and subsequently addressed through the soft actor-critic (SAC) algorithm to establish an adaptive parameter configuration pattern. With the reward function properly designed, the fine-tuned parameters under various working conditions provide the dc microgrids with optimized transient time control performance. Case studies considering various operating conditions of the dc microgrids verify its effectiveness and facility for practitioners.