【摘要】 针对传统方法下锂离子电池剩余寿命预测精度低，提出一种基于堆叠自编码器（stacked autoencoder,SAE）下集合经验模态分解（ensembleempiricalmode decomposition,EEMD）和门控循环单元神经网络（gated recurrent unit neural network,GRU）组合而成的SAEEEMD-GRU(SEG)预测方法。应用SAE深层特征表达能力，对六个电池参数进行去噪、降维，重构出一个集中包含电池退化特性的融合健康因子，并利用EEMD不平稳信号分析算法将融合健康因子进行分解，获得若干个子序列。根据GRU网络时间序列分析能力，对子序列分别建立GRU模型并叠加重构，进一步提高锂离子电池剩余寿命的预测精度。最后采用PCoE(NASA ames prognostics center of excellence)电池数据集，与SAE-GRU方法及GRU方法进行对比实验，实验结果表明了SAE-EEMD-GRU(SEG)预测方法可以有效提高锂电池剩余寿命预测精度，并使预测误差RMSE,MAE控制在2%以下。更多还原

【Abstract】 Aiming at the low accuracy of lithium-ion battery remaining life prediction under traditional methods,a SAE-EEMD-GRU(SEG)prediction method combining stacked autoencoder(SAE),ensemble empirical mode decomposition(EEMD)and gated recurrent unit neural network(GRU)is proposed.Using the deep feature expression ability of SAE,denoising and dimensionality reduction of six lithium battery parameters,reconstruction to obtain a fusion health factor that contains battery degradation characteristics,and using EEMD unsteady signal analysis algorithm to decompose the fusion health factor,obtain several subsequences.According to the GRU network time series analysis capability,GRU models are established for the sub-sequences and superimposed and reconstructed to achieve high-precision prediction of the remaining life of the lithium-ion battery.Finally,the PCoE(NASA ames prognostics center of excellence)battery data set is used to conduct comparative experiments with the SAE-GRU method and the GRU method.Experimental results show that the SAEEEMD-GRU(SEG)prediction method can effectively improve the accuracy of residual life forecasting of lithium battery and control the RMSE and MAE prediction error below 2%.更多还原