Practical Implementation Method of Reinforcement Learning for Power Converter

Reinforcement Learning (RL) is a field of machine learning that is widely used to control complex systems such as games. It has the advantage of finding optimal control input at a specific state through the iterative learning process. Recently, RL has been considered for the power electronics application having nonlinearities with multiple inputs due to empirically finding the optimal control input. This paper is about the DC/DC converter control using the DDPG (Deep Deterministic Policy Gradient). During the DDPG training, permanent damage to the experimental setup may occur. Thus, an offline training environment is created, and the neural network training for the DDPG is performed using iterative PC simulations. The acquired artificial neural network gains are directly applied to the experimental setup, and it improves the control performance compared to the conventional PI controller. Also, unlike previous papers that were performed on the costly FPGAs to solve artificial neural networks, in this paper, the size of neural networks is reduced sufficiently. Thus, the RL algorithm can be implemented with only a cost-effective microcontroller.