Concurrent charging and balancing of battery packs using deep reinforcement learning

Battery charging and balancing are critical functions of battery management systems that ensure the safety and longevity of lithium-ion battery (LiB) packs. This paper presents an adaptive deep reinforcement learning (DRL) framework for simultaneous charging and balancing of LiB packs. Traditionally, these objectives are handled sequentially through conservative, rule-based approaches that limit performance and flexibility, or through model-based approaches that typically require intensive optimization during implementation. Neither generally adapts to battery condition with age and operational conditions. In contrast, DRL can learn context-aware control policies that dynamically adjust to the evolving pack state. To support safe and sample-efficient learning, we use a hybrid virtual model that combines a physics-based battery model with neural residuals. Continuously refined with operational data, this model provides a high-fidelity environment for DRL training and avoids on-battery exploration. The task is framed as a Markov decision process in which the agent maximizes charging and balancing speed while minimizing energy loss and respecting voltage and temperature safety constraints. Compared with conventional constant current–constant voltage (CCCV) charging followed by rule-based top balancing, the proposed method achieves up to 2 × faster charging while keeping SOC imbalance below 0.25%. A balancing-only ablation of the algorithm similarly improves balancing speed by up to 2 × and reduces power-electronics losses by 50% relative to rule-based SOC balancing. The results show that learning-based strategies can adapt control policies to prevailing battery conditions and significantly outperform traditional conservative approaches.