Machine Learning–Based Control and Optimization of Power Electronic Converters

Research Article

Abstract

Power electronic converters are core enabling technologies in renewable energy systems, electric transportation, and modern power grids. Their inherent nonlinear dynamics, switching-induced harmonics, and parameter uncertainties pose persistent challenges to conventional control approaches. This paper proposes a machine learning–based control and optimization framework for power electronic converters, aimed at achieving improved dynamic performance, robustness to uncertainties, and higher operational efficiency. A deep reinforcement learning controller is developed to regulate a DC–DC boost converter, replacing conventional proportional–integral control. The proposed controller learns an optimal duty-cycle policy directly from system interaction without relying on an explicit mathematical model. A simulation-based experimental platform is established to train and test the controller under varying load and input voltage conditions. Performance is evaluated in terms of voltage regulation accuracy, transient response, efficiency, and robustness to parameter variation. Results demonstrate that the machine learning controller achieves faster settling time, lower steady-state error, and improved efficiency compared with a benchmark PI controller. Furthermore, the learned policy exhibits strong adaptability to unseen operating conditions. The study confirms the feasibility and advantages of data-driven intelligent control for next-generation power electronic converters and provides a pathway toward real-time embedded implementation.