Reinforcement Learning Strategies for Parameter Design of Bidirectional Cllc Resonant Converters With Ultrawide Voltage Range

In this paper, a design methodology of the resonant tank for a bidirectional CLLC resonant converter based on massively parallelized circuit simulations is proposed to achieve ultrawide input/output voltage range. A simplified closed-loop model of the CLLC converter with variable circuit parameters is built. To investigate potential configurations of the circuit parameters, various sampling methods and optimization techniques are used to generate diverse sets of parameters: Grid Search (GS), Tree-structured Parzen Estimator (TPE), Covariance-Matrix Adaptation Evolution Strategy (CMA-ES), standard Reinforcement Learning (RL), Diversity-driven Reinforcement Learning (DdRL). Appropriate criteria have been added to greatly narrow down the plausible sampled design configurations which can be used for further detailed analysis. Together with the conventional design method following the design guideline, the design performances are analyzed and compared. The sampling results indicate the RL-based algorithms can obtain the comparable number of effective designs as GS and have much better diversity at the same time, which demonstrates the RL-based design methods can effectively help power electronic engineers design the resonant tank parameters of the CLLC converter.