Equivalent circuit simulated deep network architecture and transfer learning for remaining useful life prediction of lithium-ion batteries

Abstract

Conventional approaches for predicting remaining useful life (RUL) of lithium-ion batteries (LiBs) have limited applicability due to their requirements of large training data. We address this issue by proposing a novel framework comprising of a feature construction technique using the charging voltages of a battery and a transfer learning architecture to prognosticate the state-of-health (SoH) of the LiB. The transfer learning approach uses a deep neural network architecture that combines equivalent circuit simulated (ECS) layers and a fine-tuning network hierarchy. The ECS-layers model the electrical equivalent circuit of the LiB converting extracted informative features to the ohmic resistance parameters proportional to a LiB's SoH. The fine-tuning architecture constructed by stacking the long short-term memory (LSTM), dropout, fully connected and regression layers determines the changes in the ohmic resistance during battery operation. The predictive performance of the proposed framework is enhanced via transfer learning. The comparison between the proposed framework and existing state-of-the-art models based on multiple battery datasets shows its better predictive performance, particularly, when the training data are sparse. The applicative example demonstrates that the proposed modeling framework allows more accurate prediction of actual degradation processes of LiBs before its end-of-life state.