Reinforcement Learning Hybrid Model for Adaptive Control of Power Electronic Converters

Abstract

The Adaptive Control of Power Electronic Converters has garnered considerable interest for mitigating uncertainties and non-idealities in contemporary systems. This work examines the viability and efficacy of an innovative adaptive control technique for power electronic converters, integrating parasitic factors to accurately mimic Boost and Buck converter configurations. An adaptive control methodology based on Lyapunov is formulated to guarantee system stability using adaptive stabilisation principles. The proposed methodology incorporates many filtering approaches, such as Median, Simple Moving Average, Hampel, and Kalman filters, to improve signal processing. Moreover, feature engineering and machine learning techniques are utilised for feature extraction and selection in classification models. A hybrid approach integrating reinforcement learning (RL) with queuing models is presented, wherein RL is learned offline utilising gathered data, while a queuing policy regulates real-time system operations, thus circumventing the inefficiencies associated with online training. The adaptive controller efficiently manages both converter topologies in non-ideal settings, while Hybrid RL substitutes traditional lookup tables with nonlinear approximators like multi-layer perceptrons, facilitating scalability to larger state spaces. The results indicate that the hybrid RL framework surpasses baseline model-based policies in both open- and closed-loop scenarios, with an accuracy of 96.21%, hence underscoring its resilience and practical application.