Health State Prediction of Lithium Ion Battery Based on Deep Learning Method

Abstract

To predict the health status of lithium-ion batteries, long and short-term memory (LSTM) recurrent neural networks are used to build two types of battery SOH evaluation models. The discharge capacity is as input to a single feature model. While, the charge capacity, the charge time, the average charge temperature, the charge average voltage, the discharge temperature, and the discharge average voltage are as input to a multiple feature input model. The data samples are separated to training and test dataset. The test results show that the maximum absolute error of the LSTM-based model is less than 2%, which satisfies the industry standard (less than 5%). Meanwhile, this study transfers the NASA-based lithium-ion battery model to the University of Maryland lithium-ion battery data, which can reduce model training iterations and get good performance as well. The experimental results validate the effectiveness of transfer learning in the field of lithium-ion battery SOH prediction and provide a reference for future work in this field.