Physics Informed Neural Network-Estimated Circuit Parameter Adaptive Modulation of DAB

• Published in: IEEE transactions on power electronics Vol. 40; no. 10; pp. 14821 - 14841
• Main Authors: Dey, Saikat, Mallik, Ayan
• Format: Journal Article
• Language: English
• Published: IEEE 01.10.2025
• Subjects: Analytical models; Artificial neural networks; Bridge circuits; Computational modeling; Dual-active-bridge (DAB); efficiency optimization; Hardware; Inductance; Integrated circuit modeling; Mathematical models; Modulation; parameter estimation; physics informed neural network (PINN); triple-phase-shift modulation (TPS); Voltage control; zero-voltage-switching (ZVS)
• ISSN: 0885-8993; 1941-0107
• DOI: 10.1109/TPEL.2025.3574873

This article presents the development, implementation, and validation of a loss-optimized and circuit parameter-sensitive triple-phase-shift (TPS) modulation scheme for a dual-active-bridge (DAB) dc-dc converter. The proposed approach dynamically adjusts control parameters based on circuit parameters estimated using a physics-informed neural network (PINN). For this purpose, first, a precise frequency-domain model for the DAB converter is developed, incorporating resistive and capacitive nonidealities, deadtime effects, and switching network losses. This model closely mimics real hardware behavior with a power flow absolute error margin of within 5%. Secondly, such an analytical model is utilized to synthesize overall system loss optimized TPS control variables as polynomial fitted functions of the converter's output voltage, power, and parameters that are subject to uncertainty such as power transfer inductances and device on-state resistances. Following this, combining a data-driven neural net (NN) and analytical DAB physics model, a PINN architecture is constructed with the purpose of estimating DAB circuit parameters with inputs of dc side voltages and currents, and modulation variables. Notably, the data-light PINN is trained with a mixed dataset collected from real hardware and the analytical DAB model. Furthermore, the trained NN is implemented in a Python environment, with serial peripheral interface (SPI) communication established with the DAB controller enabling real-time updates of the TPS modulation variables based on NN-predicted circuit parameters. The proposed NN-in-loop, parameter-sensitive TPS modulation is validated and benchmarked in a 2 kW/100 kHz GaN-based DAB converter. With an experimentally emulated 30% reduction in series inductance, the converter achieves a peak efficiency of 98.3%, representing an average efficiency improvement of approximately 1.4% compared to conventional nonadaptive TPS modulation.