Deep Deterministic Gradient Policy (DDGP) Reinforcement Learning Assisted Degradation-Aware Control of Solid-State Transformer

Abstract

The reliability of semiconductor switches is significant in power electronics systems. It is especially critical with solid-state transformer (SST) due to its safety- and mission-critical applications, including in-vehicle charging or power-grid interfaces. The traditional controller commonly requires an accurate mathematical model. However, such a traditional setting changes due to degradation and aging, adding extreme complexity and large uncertainty in a stability study. To address such uncertainty in increasingly complicated SST operations, in this paper, a deep deterministic gradient policy (DDGP) reinforcement learning (RL) assisted degradation-aware control of SST is proposed. The proposed controller will avoid complex mathematical modeling while ensuring optimal power transfer with an extended lifetime. The proposed actor-critic DDGP assisted controller learns and optimizes the phase-shift angle by evaluating the SST behavior under different input and health status. In this paper, an analytical background of DDGP and its application in a reliability integrated controller is provided along with a training environment. The validity of the proposed controller is validated by 5kW cascode GaN FET-based SST prototype.