Multi-Agent Reinforcement Learning-Based Decentralized Controller for Battery Modular Multilevel Inverter Systems

Abstract

The battery-based multilevel inverter has grown in popularity due to its ability to boost a system’s safety while increasing the effective battery life. Nevertheless, the system’s high degree of freedom, induced by a large number of switches, provides difficulties. In the past, central computation systems that needed extensive communication between the master and the slave module on each cell were presented as a solution for running such a system. However, because of the enormous number of slaves, the bus system created a bottleneck during operation. As an alternative to conventional multilevel inverter systems, which rely on a master–slave architecture for communication, decentralized controllers represent a feasible solution for communication capacity constraints. These controllers operate autonomously, depending on local measurements and decision-making. With this approach, it is possible to reduce the load on the bus system by approximately 90 percent and to enable a balanced state of charge throughout the system with an absolute maximum standard deviation of 1.1×10−5. This strategy results in a more reliable and versatile multilevel inverter system, while the load on the bus system is reduced and more precise switching instructions are enabled.