Data-Driven Surrogate Model for Voltage Control in Multi-active Bridge Converters

Multi-Active Bridge (MAB) converters offer a compact interface for interconnecting multiple voltage levels while being an isolated DC-DC converter. However, the presence of a high-frequency transformer comes at the cost of a coupled and non-linear model. To mitigate the complex structure of the model and achieve precise voltage control, this paper proposes a machine learning (ML) based control strategy for an MAB converter. The dataset required to train the ML model is obtained by applying programmed step changes in the load and input voltage, while varying the phase-shifts of the MAB converter. During this process, the port voltages together with the corresponding control commands are continuously recorded to form the training dataset. After training, the ML-based controller directly maps the voltage reference signals to optimal phase-shift commands that minimize the steady-state error between the measured and reference voltage. Consequently, the proposed approach eliminates the need for complex multi-variable controller tuning and decoupling-based control design. Experimental results are curried out to confirm the effectiveness of the proposed ML-based controller across a wide range of operating conditions and voltage regulation scenarios for the MAB converter.