Abstract:

Accurate prediction of the remaining useful life (RUL) of power batteries can avoid the risk of battery overuse, inform decision making on the secondary use of retired batteries, and improve the utilization rate of second-life batteries. We propose a method based on ensemble empirical mode decomposition (EEMD), gated recurrent units (GRUs), and temporal recurrent unit networks to reduce the dependence of RUL prediction accuracy on nonlinear features (this dependence is caused by noise and capacity recovery in the power battery RUL prediction task). GRU and temporal convolutional networks (TCNs) are integrated into the RUL prediction model for power batteries. First, the raw data are decomposed using EEMD, and the nonlinear features caused by noise and capacity rebound during power battery capacity decline are decomposed into high-frequency components. The main trends of the raw capacity data are decomposed into low-frequency components. Next, GRUs and TCNs are used to predict the high- and low-frequency components, respectively. Finally, the predictions are integrated using attention. The experimental results on the NASA dataset show that the prediction accuracy and the fitting of nonlinear features of the integrated model proposed in this paper are better than those of other single models and other models of the same type, with the maximum average absolute error and the maximum root-mean-square error within 0.52% and 0.74%, respectively, and the absolute error within one cycle period. These results prove that the proposed model produced more accurate predictions of the RUL than conventional models.

Key words: power battery, remaining service life, empirical mode decomposition, gated recurrent unit network, temporal convolutional networks