This study aims to enhance the efficiency and dynamic performance of hybrid resonant and nonresonant DC–DC converters used in photovoltaic (PV) microinverters (MIs). By addressing the challenge of real-time power path control under varying solar input conditions, the work seeks to enable intelligent and adaptive control. This approach optimizes power distribution while ensuring soft-switching and reduced losses.
A physics-guided machine learning (PGML) model is developed using time-domain simulation data from a hybrid current-fed converter. The model is trained to predict optimal power split ratios based on real-time operating conditions such as input voltage, load, switching frequency and phase shift. The control strategy minimizes conduction losses, turn-off current and power backflow while meeting design constraints.
The machine learning-assisted control approach delivers up to 5.5% higher efficiency and reduces switching stress compared to traditional static control methods. It ensures optimal soft-switching across varying PV conditions. At the same time, it eliminates the need for computationally intensive real-time optimization or lookup tables, improving response time and overall system performance.
This work pioneers the integration of physics-informed machine learning into hybrid power converter control, offering a scalable and adaptive solution for next-generation PV systems. Unlike conventional lookup table-based or heuristic control methods, the proposed PGML approach achieves sub-35 µs real-time inference on embedded hardware while maintaining high zero-voltage switching compliance (>97%) across the tested operating envelope and low current ripple. It enables up to 5.5% higher conversion efficiency and improved dynamic response across a wide input-voltage and load range. This demonstrates a scalable, adaptive control paradigm for next-generation PV MIs.
