| 摘 要: 锂离子电池的健康状态（state of health, SOH）与剩余使用寿命（remaining useful life, RUL）精准预测是电池管理系统（battery management system, BMS）的核心功能。针对电池容量退化过程中的高度非线性、长期依赖性及局部容量再生现象带来的不确定性挑战，本文提出一种融合经验模态分解（empirical mode decomposition, EMD）、长短期记忆网络（long short-term memory, LSTM）与高斯过程回归（gaussian process regression, GPR）的创新数据驱动框架。该方法通过EMD将原始容量序列自适应分解为表征局部波动的高频本征模态函数（intrinsic mode functions, IMFs）和表征长期退化趋势的低频残差(residual)分量。LSTM网络被用于捕捉残差中的长期依赖关系，而GPR则专门拟合IMFs，以量化由容量再生引起的不确定性。最终通过分量融合实现全概率预测。在美国国家航空航天局（national aeronautics and space administration, NASA）及美国马里兰大学能源研究中心（center for advanced life cycle engineering, CALCE）公开数据集上的实验表明，所提模型在1步预测中的均方根误差（root mean square error, RMSE）低至0.0032，显著优于单一模型及其他混合模型。多步及早期RUL预测结果进一步验证了该模型在精度（最大误差<1.8%）、鲁棒性及不确定性量化（95%置信区间有效覆盖真实值）方面的卓越性能，为电池健康管理提供了可靠的理论与工程实践工具。 |
| 关键词: 锂离子电池 健康状态预测 剩余使用寿命 经验模态分解 长短期记忆网络 高斯过程回归 不确定性量化 |
|
中图分类号:
文献标识码:
|
| 基金项目: 河南省高等学校重点科研项目：基于多模态信号分解与不确定性量化技术的锂电池健康状态预测模型研究 |
|
| Data-driven and uncertainty quantification-based prediction of the health status and remaining useful life of lithium-ion batteries |
|
lijiangjiang1, fenglijuan2, songqingmin2, songxianan2
|
1.Zhengzhou University of Science and Technology;2.School of Electronics and Electrical Engineering, Zhengzhou University of Science and Technology,
|
| Abstract: Accurate prediction of the health status (SOH) and remaining useful life (RUL) of lithium-ion batteries is a core function of the battery management system (BMS). In response to the uncertainty challenges brought about by the highly nonlinear, long-term dependent and local capacity regeneration phenomena during the battery capacity degradation process, this paper proposes an innovative data-driven framework that integrates Empirical Mode Decomposition (EMD), Long short-term memory network (LSTM) and Gaussian Process Regression (GPR). This method adaptively decomposes the original capacity sequence into high-frequency intrinsic mode functions (IMFs) characterizing local fluctuations and low-frequency Residual components characterizing long-term degradation trends through EMD. LSTM networks are used to capture long-term dependencies in residuals, while GPR is specifically designed to fit IMFs to quantify the uncertainties caused by capacity regeneration. Ultimately, full-probability prediction is achieved through component fusion. Experiments on the public datasets of NASA and CALCE show that the root mean square error (RMSE) of the proposed model in one-step prediction is as low as 0.0032, which is significantly better than that of single models and other hybrid models. The multi-step and early RUL prediction results further verified the model's outstanding performance in terms of accuracy (maximum error <1.8%), robustness, and uncertainty quantification (95% confidence intervals effectively cover true values), providing a reliable theoretical and engineering practice tool for battery health management. |
| Keywords: Lithium-ion battery Health status prediction Remaining service life Empirical mode decomposition Long short-term memory network Gauss process regression Quantification of uncertainty |
