Enabling Safety-Enhanced fast charging of electric vehicles via soft actor Critic-Lagrange DRL algorithm in a Cyber-Physical system

Abstract

Fast charging of lithium-ion battery (LIB) is critical for the further popularity of electric vehicles. However, the partial pursuit of high-power charging can violate the physical limits of LIB, and further cause unexpected side effects or even catastrophic safety issues. Motivated by this, this paper proposes a multi-state-constrained fast charging strategy for LIB, enabled by a novel deep reinforcement learning (DRL) technique. In particular, the SAC-Lagrange DRL is developed, for the first time, to train the fast-charging strategy assisted by an electro-thermal model for unmeasurable state perceiving. The proposed strategy is further performed within a cyber-physical system-based management framework, where the complicated training is carried out in the cloud, while the trained low-complexity policy is executed in the onboard controller to mitigate the risk of high computing burden. Hardware-in-Loop tests and practical charging experiments are carried out for validation. Results reveal that the proposed strategy can boost the charging speed remarkably while smartly avoid violating the electrical and thermal safety limits of LIB. Compared with the advanced DDPG-based and SAC-based strategies, the proposed one has improved the optimality and stability, and thus enjoys more guaranteed convenience and versatility in practical charging applications.