Adaptive Transfer Learning Strategy for Predicting Battery Aging in Electric Vehicles

This work presents an adaptive transfer learning approach for predicting the aging of lithium-ion batteries (LiBs) in electric vehicles using capacity fade as the metric for the battery state of health. The proposed approach includes a similarity-based and adaptive strategy in which selected data from an original dataset are transferred to a clean dataset based on the combined/weighted similarity contribution of feature and stress factor similarities and times series similarities. Transfer learning (TL) is then performed by pre-training a model with clean data, with frozen weights and biases to the hidden layer. At the same time, weights and biases toward the output node are recalculated with the target data. The error reduction lies between −0.4% and −8.3% for 20 computational experiments, attesting to the effectiveness and robustness of our adaptive TL approach. Considerations for data structure and representation learning are presented, as well as a workflow to enhance the application of transfer learning for predicting aging in LiBs.