A Reinforcement Learning-based Online-training AI Controller for DC-DC Switching Converters

A controller for DC-DC switching converters based on solely AI algorithm is proposed with a simpler structure than the traditional neural network-PID controllers. Reinforcement learning is used to train the AI controller online using deep deterministic policy gradient (DDPG) algorithm. The AI controller with an actor-critical architecture realizes model-free control with strong self-adaptive ability for different control objects, which can be used for different types of DC-DC switching converters. The performance of a buck DC-DC switching converter with the AI controller is compared with a neural network-PID controller through simulation. The simulation results show that the settling time is improved by at least 65% and overshoot/undershoot is decreased by at least 43%.

I. Introduction

DC-DC switching converters are widely used in portable electronic devices, electric vehicles, and the Internet of Things applications (IoT) [1] –[4]. In recent years, in order to further improve the performance of digital power supplies and the adaptability of constant power loads (CPLs), intelligent digital controllers based on a neural network (NN) have attracted a lot of attention [5]. Compared with the traditional control methods, the NN-based controllers have the feature of faster response speed, wider adaptability and better robustness.