A multi-agent reinforcement learning approach for continuous battery cell-level balancing

Cell balancing in lithium-ion battery packs is essential for optimal performance, with active balancing offering advantages over passive balancing but requiring complex management that simple methods fail to address. This paper proposes a multi-agent reinforcement learning (MARL) approach for continuous cell-level balancing, particularly for electric vehicles (EVs). By treating each cell as an independent agent, MARL enables decentralized, cooperative management and goal-oriented decision-making, overcoming the limitations of centralized and rule-based methods. The intelligent MARL approach manages the flow of charge between cells, utilizing a shared low-voltage bus and individual DC-DC converters to maximize pack capacity and minimize state-of-charge (SoC) variance. Trained with the trust region policy optimization (TRPO) algorithm, the approach ensures stability and partial observability. Extensive simulations demonstrate the MARL model's superiority over baselines, showing a 28% improvement in battery performance and balancing speed up to three times faster under EV load profiles. This approach significantly enhances battery efficiency, contributing to longer-lasting, sustainable solutions through intelligent decentralized management. • Novel application of multi-agent reinforcement learning for Li-ion battery cell balancing. • Decentralized, agent-based approach: cells manage their own DC-DC converters. • Flexible, goal-oriented model for complex balancing objectives beyond SoC unification. • 28% improvement in battery performance and balancing speed up to three times faster. • Potential for enhanced battery management in electric vehicles and beyond.