Safety-Optimized Fast Charging of Lithium-Ion Battery Based on Distributional SAC-Conservative Augmented Lagrangian SDRL Algorithm

The widespread utilization of electric vehicles depends on the advancement and optimization of fast battery charging technology. To tackle the dual challenges of attaining rapid charging while guaranteeing safety, a novel safety-enhanced fast charging strategy for lithium-ion batteries (LIBs) is presented, utilizing a thermal and health-aware safe deep reinforcement learning (SDRL) approach. Specifically, this work introduces an innovative approach by employing distributional reinforcement learning in LIBs fast charging control, simultaneously optimizing safety constraint issues. The proposed distributional soft actor critic-conservative augmented Lagrangian (DSAC-CAL) algorithm mitigates the overestimation of the reward value function while avoiding the underestimation of the cost value function. Both simulation and real-world charging experiments are conducted to validate the strategy's effectiveness. Compared to advanced deep deterministic policy gradient-based and soft actor-critic-based strategies, the proposed method exhibits superior optimality and stability, confirming its significant performance advantages. Furthermore, in contrast to model predictive control-based charging strategies, it offers superior real-time adaptability and proves more efficient in practical charging scenarios.