Deep Reinforcement Learning and Ultra-Local Model-Free Hybrid Control Method for Dual Active Bridge Converter

To address the challenges of controlling a dual active bridge (DAB) converter under nonlinear loads that cannot be accurately modeled, this paper proposes a hybrid control method combining deep reinforcement learning (DRL) and ultra-local model-free control (ULMC). The proposed approach eliminates the need for precise mathematical models, ensures accurate regulation of the DAB under nonlinear loads and large disturbances, and enhances system adaptability. First, a model free Kalman filter (MFKF) is developed to filter out noise and estimate uncertain disturbances. Next, the ULMC is designed to ensure system stability without relying on an accurate model of the system, even under large-signal disturbances. Finally, a DRL-based sub-controller using proximal policy optimization (PPO) is designed, which is capable of learning robust and risk-sensitive control policies under model uncertainty and large disturbances. Moreover, the PPO-based sub-controller is integrated into the ULMC to intelligently adjust the control parameter, further improving adaptability and performance. The experimental results demonstrate that the proposed control method can adaptively maintain the stability of the system under large signal disturbance without requiring an accurate model, and achieve excellent dynamic and steady-state performance.