A hierarchical deep reinforcement learning-model predictive control framework for lithium-ion battery fast charging under expansion-force constraints

Fast charging of lithium-ion batteries is constrained by the trade-off between charging efficiency and physical safety, especially when high-current operation narrows the admissible electrical, thermal, and mechanical margins. Traditional model-based controllers can explicitly handle constraints, but their performance depends on model fidelity and may become conservative near safety boundaries. In contrast, model-free deep reinforcement learning (DRL) provides flexible sequential decision-making but does not by itself guarantee hard safety-constraint satisfaction. This study presents a multi-stage constant-current (MCC)-inspired hierarchical DRL–model predictive control (MPC) framework that couples staged-current policy search with MPC-based safety correction under expansion-force constraints. Expansion force is incorporated as a dynamic safety constraint together with terminal voltage and surface temperature, allowing their coupled responses to be considered during control. A long short-term memory (LSTM)-based surrogate model is developed to predict battery-state evolution and support control-oriented optimization. Independent trajectory- and C-rate-holdout tests are used to examine its multi-step prediction error. Based on the surrogate predictions, the upper DRL layer selects candidate staged-current actions, whereas the lower MPC layer performs quadratic-programming (QP)-based safety correction before execution. Comparative simulation results against 3C constant-current (CC), native MPC, and pure DRL baselines show that the proposed method reaches 90% state of charge (SOC) in 728 s and improves the comprehensive score by 23.02%, 6.53%, and 11.64%, respectively. The reported trajectories remain below the prescribed voltage and temperature limits and the MPC expansion-force operating envelope.