一种基于模态分解和机器学习的锂电池寿命预测方法

A lithium battery life-prediction method based on mode decomposition and machine learning

锂离子电池剩余使用寿命(RUL)是电池健康管理的一个重要指标。本工作采用电池容量作为健康状况的指标,使用模态分解和机器学习算法,提出了一种CEEMDAN-RF-SED-LSTM方法去预测锂电池RUL。首先采用CEEMDAN分解电池容量数据,为了避免波动分量里的噪音对模型预测能力的影响,且又不完全抛弃波动分量里的特征信息,本工作提出使用随机森林(RF)算法得到每个波动分量的重要性排序和数值,以此作为每个分量对原始数据解释能力的权重。然后将权重值和不同波动分量构建的神经网络模型得到的预测结果进行加权重构,进而得到锂离子电池的RUL预测。文章对比了单一模型和组合模型预测精度,加入了RF的组合模型预测精度让五种神经网络的表现都有进一步的提升。最后,对表现较好的两种网络——LSTM和GRU引入了简单编码解码(SED)的机制,让其更好地学习到序列数据全局时间上的特征和远程的依赖关系。以NASA数据集作为研究对象进行该方法的性能测试。实验结果表明,CEEMDAN-RF-SED-LSTM模型对电池RUL预测表现效果好,预测结果相比单一模型具有更低的误差。

Lithium-ion battery’s remaining useful life(RUL)is an important indicator of battery health management.Therefore,in this paper,using battery capacity as an indicator of health status,including modal decomposition and machine learning algorithms,a CEEMDAN-RF-SEDLSTM method was proposed to predict lithium battery RUL.First,adaptive white-noise fullensemble empirical-mode decomposition(CEEMDAN)was used to decompose the battery capacity data.Then,to avoid the influence of the noise in the fluctuation component on the prediction ability of the model and not completely discard the characteristic information in the fluctuation component,this paper used the Random Forest algorithm to obtain important values for each fluctuation component,after which sexual ranking and numerical values were used as weights for each component’s ability to explain the original data.Subsequently,the weight value and prediction result obtained by the neural network model constructed by different fluctuation components were weighted and reconstructed,resulting in the RUL prediction of the lithium-ion battery.Next,this research compared the prediction accuracy of the single model and the combined model,followed by the addition of the combined model prediction accuracy of RF,to improve the performance of the five neural networks further,after which the Simple Encoder-Decoder(SED)mechanism was introduced for the two networks with better performance,LSTM and GRU,to better learn the global temporal features and long-range dependencies of sequence data.We finally tested the method’s performance using the NASA dataset as the research object.The experimental results showed that although the CEEMDAN-RF-SED-LSTM model performed well in battery RUL prediction,the prediction results had lower errors than the single model.