Artificial-Neural-Network Optimized Triple-Active-Bridge Converter Minimizing Total Semiconductor Losses

A data-driven modulation strategy for a triple-active-bridge (TAB) DC-DC converter is developed to directly minimize the total semiconductor losses, including both conduction and switching components, in wide-bandgap (WBG) SiC MOSFET-based systems. An offline exhaustive parameter sweep in PLECS, integrating detailed device-level SiC loss models and enforcing zero-voltage switching (ZVS) constraints, generates a high-fidelity dataset. The resulting dataset is further augmented and used to train a compact, artificial neural network (ANN) that predicts loss-optimal duty-cycles in real time, eliminating the need for large multidimensional lookup tables and computationally intensive online searches. Simulation results demonstrate consistent loss reductions of up to 23% compared to conventional phase-shift modulation across a wide operating envelope spanning 200-1000V and 2-10kW. The trained ANN requires only 870 bytes of memory and achieves mean absolute prediction error below 0.8%, enabling practical, high-efficiency control of multiport converters with minimal computational overhead and facilitating scalable implementation in embedded real-time power electronic controllers.