A learning-based method for battery health balancing and capacity maximization in reconfigurable battery systems

Lithium-ion batteries are increasingly deployed in large-scale energy storage and electric mobility systems, where performance degradation and cell inconsistency pose critical challenges at the system level. Among these challenges, achieving effective State-of-Health (SOH) balancing remains particularly difficult, especially in dynamic reconfigurable battery (DRB) systems. To address the absence of SOH-oriented balancing strategies in DRB architectures, this work introduces a latent metric, termed the Relative State of Charge (ReSOC), and proposes an ReSOC-based deep reinforcement learning algorithm, referred to as R-M-DQN, to simultaneously enhance SOH uniformity and maximize system-level energy utilization. Comprehensive case studies demonstrate that the proposed R-M-DQN approach substantially improves SOH balancing performance and maximizes deliverable system capacity, outperforming state-of-the-art balancing strategies.