Parameter identification of fractional-order model with transfer learning for aging lithium-ion batteries

The aging of lithium-ion batteries (LiBs) is inevitable during their operation owing to their irreversible side reactions. It is practical to capture only the dominant physicochemical processes with a physics-based model for engineering applications, as the degradation mechanism of LiBs is complex and interconnected. Numerous factors dramatically affect the performance of LiBs; thus, it is necessary to use the real-time operational data to estimate the model parameters online. The combination of the degradation mechanism and real-time data can effectively reduce the complexity of the parameter identification, and offers practicality from the perspective of engineering applications. In this study, a transfer learning method based on a back propagation neural network (BPNN) is introduced as the data-driven parameter identification method for the physics-based fractional-order model (FOM) of the LiBs. Accelerated aging tests are designed for commercial LiBs to reduce the experimental time, and the reference performance tests under different aging states are implemented to capture the degradation modes with increasing cycle numbers. The initial key parameters of the fresh FOM are estimated based on the BPNN with two hidden layers. The model-based transfer learning is proposed to estimate the key parameters of the aging FOM. The parameter identification results at different aging states are validated under various working loads. The good fit with the experimental data indicated that the parameters can be identified accurately using the transfer learning with BPNN. The outstanding results show the effectiveness of the parameter identification method based on transfer learning.

Automated summary

This study presents a transfer learning method based on a back propagation neural network (BPNN) as a data-driven parameter identification method for LiBs. The combination of degradation mechanism and real-time data effectively reduces parameter identification complexity, with experiments demonstrating accurate parameter identification using transfer learning with BPNN.