Optimal Design of Planar Transformer for DAB Converters Based on Model-Free Reinforcement Learning

The traditional high-frequency transformer used in Dual Active Bridge (DAB) converters often employs Litz wire, which poses manufacturing difficulties and necessitates the inclusion of additional inductors. To mitigate these problems, this paper presents a planar transformer design with PCB-based coils that takes advantage of the transformer’s natural leakage inductance, thereby eliminating the need for extra inductors. Additionally, a design optimization approach using reinforcement learning is introduced. The optimization process, implemented with the Deep Q-Network (DQN) algorithm in Python and TensorFlow, is aimed at determining the ideal coil widths on both the primary and secondary sides of the PCB to reduce copper losses. The proposed method is validated through simulations running at 100 kHz, with the performance assessed over 200 episodes, each consisting of 50 steps.