Deep Reinforcement Learning-Based Intelligent Battery Charging Strategy

Abstract

To address the conflict between charging speed, safety, and lifespan in lithium-ion batteries, this paper proposes a physics-consistent and deployable deep reinforcement learning (DRL) intelligent charging strategy. Existing DRL studies often overlook the algebraic loop coupling between power and voltage, leading to nonphysical oscillations in control signals. To overcome this, we construct a high-fidelity digital twin environment with a closed-loop current analytical solution, ensuring physical authenticity during training. Building on this, an improved Soft Actor-Critic (SAC) framework is proposed, featuring a multidimensional normalized reward function for multi-objective optimization and integrating Neural Architecture Search (NAS), pruning, and INT8 quantization for model lightweighting. Experiments across four representative scenarios (standard, low-temperature, high-temperature, and aged battery conditions) demonstrate that, compared to baselines like CC–CV and DDPG, the proposed strategy reduces charging time by 12.5% while effectively regulating peak temperature within scenario-adaptive safety limits and eliminating sawtooth oscillations. Statistical validation (p < 0.01) across 200 independent evaluation runs confirms the significance of the reported improvements. Furthermore, robustness analysis under practical uncertainties including sensor noise and control latency demonstrates the policy's resilience for real-world deployment, and resource evaluation confirms its feasibility for embedded Battery Management Systems (BMS).