Active Balancing of Reconfigurable Batteries Using Reinforcement Learning Algorithms

Paper i proceeding, 2023

In reconfigurable batteries, series or parallel connections among cells/modules are able to be actively changed during operations. One big advantage of reconfiguration is to achieve active balancing among cells/modules. Rule-based and greedy algorithms of reconfigurable battery control have problems of being sensitive to battery characteristic variation and requiring a lot of computing resources. Therefore, deep reinforcement learning (DRL) algorithms are used to overcome these difficulties. Very few studies related to this idea have been done previously, and the studied battery reconfiguration topologies are either too simple or too complex. Thus, in this paper, a module-level reconfigurable battery with moderate flexibilities is controlled by DRL algorithms. Two neighboring modules are connected in either parallel or series by following a well-trained optimal policy. Two battery discharging cases, constant power and variable power, are simulated. The final results prove the feasibility and great potential of utilizing DRL algorithms in reconfigurable battery control.