Improving Aging Identifiability of Lithium-Ion Batteries Using Deep Reinforcement Learning

As lithium-ion batteries age, the lithium inventory and active materials are gradually lost, limiting their lifespan. The stoichiometric range, which refers to the operable range of the amount of lithium in the electrode, has been considered a representative and comprehensive indicator for predicting the aging process. For the efficient and safe use of lithium-ion batteries, the cathode and anode stoichiometric ranges should be identified as accurately as possible. Accordingly, because the identification accuracy depends on the input signals and system operating conditions, suitable input current profiles should be designed for various operating conditions to improve identifiability. This paper proposes a deep reinforcement learning-based identifiability improvement scheme to estimate the stoichiometric range of a lithium-ion battery more accurately. In particular, a well-known reinforcement learning scheme (i.e., twin delayed deep deterministic policy gradient) is employed with an inverted bottleneck network identifier. The policy determines a suitable current input profile every second by considering previous voltage and current profiles. The simulation results show that the proposed scheme can provide an identifiability-improved current input profile, even under different initial state-of-charge conditions. Experiments with fresh and aged batteries were conducted to validate the proposed scheme. IEEE