Life prediction model for lithium-ion battery via a 3D convolutional network enhanced by channel attention considering charging and discharging process

Abstract

While lithium batteries provide high efficiency and low cost, accurately predicting the cycle life of batteries under different charging protocols remains a challenge. The usage of batteries with inadequate cycle life can potentially introduce safety hazards. In this study, a Depthwise Separable 3D Convolutional Network Model Fusing Channel Attention (DS-3DCA-CNN) model considering charging and discharging process is proposed for life prediction of lithium batteries. Firstly, the recurrence plot is used to transform varied cycle charging data into multidimensional form, simultaneously extracting relevant features from discharging data and analyzing their correlation with battery cycle life. Secondly, the Depthwise Separable 3D convolution is used for quicker model training with fewer parameter calculations and introduce a 3D Channel Attention (3DCA) module to increase channel interactions while keeping model complexity low. Finally, ablation experiments are conducted to explore the influence of different time series imaging methods on the accuracy of model prediction results. Experimental results reveal that the proposed DS-3DCA-CNN model, using only 10 initial cycles, predicts battery cycle life with an average error of 35 cycles and achieves a 16-cycle average error when predicting remaining useful life with 20 window cycles of data.