Multi-Agent Learning-based Control of Hybrid Battery Management Systems

Abstract

The proliferation of electrically powered vehicles has not proceeded as quickly as climate pacts aiming to reduce greenhouse gas emission would require it. Alongside improvements in battery technology, optimizing the management of battery systems can empower the switch from internal combustion engines to electric vehicles. As the performance of battery systems can be limited by the weakest individual battery cells, maintaining balance among the cells is a primary goal of battery management systems. The smart hybrid battery system developed at the Institute of System Dynamics and Control at the German Aerospace Center (DLR) augments a traditional battery storage system with dedicated power converters and supercapacitors, allowing it to redistribute energy between battery cells. Through intelligent control, this system is capable of balancing cell states efficiently and protecting weak cells from aggressive use. In this thesis, the task of generating the setpoints of the balancing currents is solved by a decentralized multi-agent reinforcement learning approach. Each of the battery modules is afforded a separate control element, which is aimed at providing scalability to battery systems of realistic sizes.