Microcontroller-Based Reinforcement Learning Control for a ZETA Converter
Felix Ecker, Nils Szabó, Julian Feth, Max Wagner, Kai Franck, Rodrigo Coelho, Christian Schumann, Andreas Rosskopf · 2025
The control of higher-order DC/DC converters such as the ZETA converter over a wide operating range is challenging with classical methods due to nonlinearities and complex component interactions. Reinforcement learning combined with neural networks provides a framework for hardware-aware training across the full operating range through iterative interaction with the system. This improves the efficiency of the control structure and its responsiveness to dynamic system behavior in power converters. This work presents a microcontroller-based reinforcement learning approach for optimizing two neural-network-based controlling strategies for a ZETA converter. The initial neural networks are generated via supervised learning on simulation datasets of the converter, tailored to the two control strategies. In the first approach, the neural network itself serves as the controller by directly computing the duty cycle. In the second approach, a sliding-mode controller is employed, where the neural network determines the control parameters.The results demonstrate that the proposed gradient-free reinforcement learning method yields strong optimization performance for the second approach. In contrast, the first approach requires further improvements, particularly in terms of network size and architecture.