Exploring Reinforcement Learning Algorithms for Current Control of Single-Phase AC/DC Full-Bridge Converters

Abstract

We discuss a reinforcement learning (RL) algorithm to create a general current control structure for a converter based on a Deep Neural Network (DNN). We discuss an RL algorithm which learns the design of a current controller for a SinglePhase AC/DC Full-Bridge Converter without any pre-calculated training data provided by the user and and/or knowledge about the circuit. No dependencies of the four switching signals are known to the RL algorithm, so it must learn to avoid bridge-leg shorting. The RL algorithm attempts to achieve control with a sinusoidal input current, and optimizes the DNN-based controller for minimum switching losses, minimum low-order harmonics, and low THD of the grid current. We implement this DNNbased controller in a circuit simulator, and analyze controller performance in the time domain.