| 摘 要: 锂离子电池剩余寿命(RUL)的准确预测对电池管理系统至关重要。针对长短期记忆网络(LSTM)超参数依赖人工经验、预测结果缺乏不确定性量化的问题,本文提出一种融合贝叶斯超参数优化与 Bootstrap 集成学习的 LSTM 预测方法。利用 NASA 公开老化数据集,首先对容量序列进行平滑和归一化,通过滑动窗口构造监督样本；然后采用贝叶斯优化自动搜索 LSTM 的最优超参数；最后通过 Bootstrap 重采样训练 20 个 LSTM 模型进行集成,输出容量预测的均值与 95% 置信区间,并递归估计剩余寿命。实验结果表明,在 B0006 测试电池上,容量预测的决定系数 R2 达 0.9609,均方根误差仅为 0.0421 Ah,平均绝对百分比误差为 2.19%；从第 100 循环预测剩余寿命,绝对误差仅 2 个循环,且真实值落在置信区间内。该方法实现了高精度和可量化的不确定性估计,为电池健康管理提供了可靠工具。 |
| 关键词: 锂电池 剩余寿命预测 LSTM 贝叶斯优化 Bootstrap 集成 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,项目编号:26A413009；2024年郑州地方高校战略新兴专业建设点支持项目(郑州科技学院机器人工程专业) |
|
| Remaining Life Prediction of Lithium Batteries Based on Bayesian Optimization and Bootstrap Integrated LSTM |
|
li jiangjiang, Zhao Junyan, Feng Lijuan, Wang Yixiao
|
Zhengzhou University of Science and Technology
|
| Abstract: Accurate prediction of the remaining useful life (RUL) of lithium-ion batteries is crucial for battery management systems. To address the issues of LSTM"s hyperparameters relying on manual experience and the lack of uncertainty quantification in prediction results, this paper proposes a prediction method that integrates Bayesian hyperparameter optimization and Bootstrap ensemble learning for LSTM. Using the NASA public aging dataset, the capacity sequence is first smoothed and normalized, and supervised samples are constructed through a sliding window; then, Bayesian optimization is used to automatically search for the optimal hyperparameters of LSTM; finally, 20 LSTM models are trained through Bootstrap resampling for ensemble, and the mean and 95% confidence interval of capacity prediction are output, and the RUL is recursively estimated. Experimental results show that on the B0006 test battery, the coefficient of determination R2 of capacity prediction reaches 0.9609, the root mean square error is only 0.0421 Ah, and the mean absolute percentage error is 2.19%; when predicting the RUL from the 100th cycle, the absolute error is only 2 cycles, and the true value falls within the confidence interval. This method achieves high accuracy and quantifiable uncertainty estimation, providing a reliable tool for battery health management. |
| Keywords: Lithium-ion battery Remaining useful life prediction LSTM Bayesian optimization Bootstrap ensemble Uncertainty quantification |
