用于宽负载应用的单输入多输出双向谐振变换器的物理信息神经网络建模

This article presents a physics-informed neural network (PINN) modeling approach for power converters with a high number of degrees of freedom. In contrast to traditional modulation strategies, which depend on harmonic approximations or time-domain analysis, the proposed method accounts for duty cycles, phase shifts, and input/output power relationships to help identify more efficient operating points (OPs). A resonant multioutput converter is used as a case study, featuring a five-level T-inverter on the primary side and a cascaded two-cell multilevel inverter on the secondary side. This topology maintains isolation between the high-voltage input and low-voltage outputs, and the cascaded structure reduces transformer turns, leading to possible improvements in power density. A particle swarm optimization (PSO) algorithm is applied to the PINN-predicted data to optimize performance further and identify optimal parameter combinations. A low-power prototype is implemented to validate the approach, demonstrating efficiency gains under light-load conditions. The results highlight the potential of artificial intelligence (AI)-driven modeling and optimization in extending converter efficiency across diverse operating scenarios. Tradeoffs, limitations, and future research directions are discussed, including digital twin integration and application to other topologies.