Online Estimation of Lithium-Ion Battery Capacity Using Transfer Learning

Abstract

Producing satisfactory accuracy in capacity estimation of lithium-ion (Li-ion) rechargeable batteries based on a small size of charge-discharge cycling data is a challenging task, since the cycling data may not cover high cell-to-cell variability in the aging process. However, in real-world applications, collecting long-term cycling data from a large number of cells is a costly and time-consuming process. This paper presents a transfer learning-based method for cell-level capacity assessment while only having access to a relatively small dataset. Transfer learning is a knowledge learning method that leverages knowledge learned from a source task to improve learning on a related but different target task. In this study, ten-year daily cycling data from implantable Li-ion cells is used as the source dataset to pre-train a deep convolutional neural network (DCNN). The parameters of this pre-trained DCNN are then transferred to a new DCNN model named deep convolutional neural networks-transfer leaning (DCNN-TL). The DCNN-TL model is then fine-tuned and re-trained to produce accurate capacity estimation on a target dataset (NASA data). Compared with the Gaussian process regression method and DCNN without transfer learning, the proposed DCNN-TL method is demonstrated to reduce the RMSE in capacity estimation by 63.09% and 17.57%, respectively.