A Generalizable Method for Capacity Estimation and RUL Prediction in Lithium-Ion Batteries

Data-driven methods have attracted much attention in capacity estimation and remaining useful life (RUL) prediction of lithium-ion batteries. However, existing studies rely on complex machine learning models (e.g., Gaussian process regression, neural networks, and so on.) that are applicable to specific observed operating conditions, and the prediction accuracy can be affected by different usage scenarios. This paper proposes to adopt a linear and robust machine learning technique, partial least-squares regression, for battery capacity estimation, and RUL prediction based on the partial incremental capacity curve. The features can be easily obtained by interpolation of the measured charging profiles without data smoothing, and the bootstrapping technique is used to give confidence intervals of the predictions, which helps to evaluate the robustness and reliability of the model. The proposed method is validated on three battery data sets with different operating conditions provided by NASA. We train the model on one battery and test its performance on the other two batteries without changing the model weights. Experimental results show that the suggested classical method exhibits greater generalizability compared to complex and sophisticated methods proposed in the literature.