A Deep Transfer Learning Framework for Li-ion Battery Temperature Prediction Based on LSTM Pretraining Model

Due to the promotion of electric vehicles and new energy, lithium-ion batteries (LIBs) have been widely used. However, temperature exerts a significant impact on the performance and safety of LIBs during operation. Therefore, it is very important to predict the temperature of LIBs and implement thermal warning. To address this issue, this paper designs a deep transfer learning method based on Long Short Term Memory (LSTM) pretraining fine-tuning model to predict the temperature of LIBs. First, select the input parameters according to the law of thermodynamics and pearson correlation coefficient. Secondly, use autocorrelation coefficient to analyze the data state correlation, and select the step size of the input parameters. Then, use LSTM to pretrain the model and freeze the relevant parameters. Finally, migrate the frozen model to the target domain and add a fully connected layer for fine-tuning training to predict battery temperature. To confirm the efficiency of the method, several complete experimental verifications were carried out on the data set of UDDS dynamic working conditions and temperatures from 0°C to 40°C. Experimental results demonstrate that the proposed method significantly improves the accuracy of LIBs temperature prediction compared to other popular data-driven models.