Li-Ion Battery Prognostics Using Statistical Models and RNN Trained on EIS-Based Features

This study addresses the challenge of accurate Li-ion battery state of health (SOH) and remaining useful life (RUL) prognostics, particularly for second-life forecasting. A model is developed that effectively captures complex nonlinear degradation behaviors under diverse operational conditions by combining statistical and deep learning techniques. Specifically, it integrates the computational efficiency of an auto-regressive integrated moving average (ARIMA) model with the predictive power of a recurrent neural network (RNN) using Long Short-Term Memory (LSTM) layers, achieving robust predictions with relatively low computational and data requirements. The approach leverages features derived from electrochemical impedance spectroscopy (EIS) measurements and empirical cycling data. The model was tested using experimental data from extensive cycling tests on eight NMC 18650 Li-ion cells under varied first-life and second-life conditions, generating a rich dataset for model training and evaluation. The hybrid ARIMA-BiLSTM model achieved exceptional predictive performance, with root mean squared errors consistently below 0.17% SOH and a mean absolute error as low as 0.054%. Analysis using equivalent full cycles (EFC) as an important metric for normalizing energy throughput reveals that both depth of discharge (DOD) and$\boldsymbol{C}$ratesignificantly influence Li-ion battery degradation, with their combined effects varying across operational conditions. These findings provide actionable guidelines for optimizing driving profiles to minimize battery degradation in EVs and enhance the performance and longevity of second-life applications.