Transfer learning LSTM model for battery useful capacity fade prediction

Abstract

Lithiumion (Li-ion) batteries have become increasingly useful within the automotive industry and modern life applications due to high energy and power densities. However, these batteries suffer capacity loss due to different ageing mechanisms in various applications. Despite several existing models, lack of accurate predictability of capacity degradation limits the advancement of Li-ion batteries. The present work focuses on prediction of battery useful capacity degradation using long-short term memory (LSTM) transfer learning neural network model. At first, a base model was developed and trained using all the (100%) degradation data available at 0°C and 10°C environmental temperatures. Thereafter, the training of the base model was fixed, and additional hidden layers were added on top of the base model to further fine tune it with only the initial 30% degradation data available at 25°C environmental temperature. The remaining (70%) data of the 25°C case was tested for model prediction. To decide the number of fixed hidden layers to be transferred from base model to transfer model and the number of additional hidden layers on top, an optimization for minimum cross validation error was performed. It was found that the resulting model was able to forecast the remaining battery degradation with 96% accuracy. The model prediction was also compared with LSTM deep learning architecture without using transfer learning. The LSTM with transfer learning model was found to be 17% higher in prediction accuracy than that without utilizing transfer learning.