A state of health estimation method for lithium-ion batteries based on initial charging segment and Gated Recurrent Unit neural network

Abstract

With the recent shortage of fossil energy and the escalating severity of environmental issues, electrochemical energy storage has emerged as a developing field. The widely used lithium-ion battery (LIB) is renowned for its exceptional performance. However, its safety concerns have garnered increasing attention. Accurate prediction of the state of health (SOH) of LIBs is crucial in mitigating safety accidents. In this study, the SOH of LIBs is predicted by selecting the initial charging segment data as features of a deep learning NN processed using dQ/dV. The processing results provide insights into the phase transformation process and aging information of both anode and cathode materials, which exhibit strong correlations with the aging behaviour of LIBs. Gated Recurring Unit (GRU) are then used to estimate SOH of LIBs. After applying dQ/dV processing to the data, the determination coefficients R² for complete charging segments in three datasets increase from 0.79, 0.47, and 0.83 to 0.96, 0.97, and 0.99, respectively. By replacing Long Short-Term Memory (LSTM) with GRU, R² values for the first 2 min of dataset 1 and dataset 2 improve from 0.32 to 0.37 to 0.93 and 0.80, which means that the use of GRU can substantially improve the prediction accuracy even though the data segment coverage time is short. This approach not only improves the estimation accuracy, but makes the entire work more interpretable and possible for application.