Voltage Control of a Quasi Z-Source Converter Under Constant Power Load Condition Using Reinforcement Learning

Abstract

The goal of this paper is to design an adaptive optimal controller for voltage control of DC–DC Z-Source (ZS) converters under a challenging load situation called Constant Power Load (CPL) without a need for any knowledge about their dynamics. These converters have been developed to remedy the drawbacks of the conventional converters and several works have been recently carried out to increase the voltage gain and decrease the voltage stress in these converters. Meanwhile, from a control point of view, voltage control of a very high voltage gain ZS DC–DC converter is unquestionably more sophisticated than the traditional converters (like Boost). Firstly, due to the non-minimum phase functionality, ZS converters present an additional challenge to controller design processes compared with traditional converters. Secondly, their output voltage sensitivity is high, and small oscillations in the control signals can lead to large oscillations in the output voltage. On the other hand, CPLs can further threaten the performance of the converter’s voltage controller due to incremental negative resistance. To alleviate these problems Reinforcement Learning (RL), one of the promising optimal and adaptive model-free control algorithms, can be used. Contrary to traditional controllers which manipulate PWM’s duty cycle, here the proposed controller directly controls the power switches inside the converter. Due to the fact that the action space is discrete and finite, this gives rise to significant simplifications in the architecture of the RL-based controller and better convergence in the training phase. Voltage control of various ZS converter topologies can be carried out by the proposed algorithm. In this paper, for instance, it is experimentally applied to a quasi ZS converter under CPL condition, and its performance compared to other classical controllers such as single-loop PI controller, dual-loop PI controller, and finite state model predictive controller is reported which confirms the better performance of the proposed algorithm in terms of voltage stability and transient tracking error.