计算机科学
领域(数学分析)
工程类
功率(物理)
控制工程
班级(哲学)
钥匙(锁)
电子工程
控制系统
算法
作者
Xizhe Wang,Xiaobin Song,Boman Su,Benben Jiang
标识
DOI:10.1109/tte.2026.3704333
摘要
Fast charging is increasingly recognized as a key requirement for the large-scale adoption of lithium-ion batteries (LIBs) in electric vehicles and energy systems. However, most existing optimization strategies are designed under fixed conditions or fixed battery parameters, which limits their ability to cope with the variability observed in practice. Such variability arises not only from manufacturing inconsistencies, such as differences in diffusion coefficients among cells, but also from external operating environments, making robust and adaptable charging strategies particularly challenging to develop. In this work, an uncertainty-aware active domain randomization (ADR) framework combined with deep reinforcement learning is proposed to address these challenges. The method employs a Gaussian-process surrogate to approximate the relation between environment parameters and policy performance, and utilizes the estimated uncertainty to adaptively select the most informative training scenarios. By focusing exploration on uncertain regions of the parameter space, ADR accelerates policy learning and strengthens the ability of reinforcement learning to generalize across diverse batteries and operating conditions. The framework is validated using a porous-electrode-theory-based battery simulator, demonstrating its effectiveness in learning fast-charging strategies that achieve reliable zero-shot transfer within the training domain and efficient few-shot adaptation under out-of-distribution scenarios, enabling robust application across heterogeneous cells and operating conditions.
科研通智能强力驱动
Strongly Powered by AbleSci AI