Capacity and RUL Prediction of Retired Batteries Using Machine Learning Features

Abstract

With the widespread application of lithium batteries, estimation of the capacity and remaining life of retired batteries has become an important issue. Traditional battery calibration tests bring high cost, and cannot estimate health status of batteries in time. For life prediction problems, retired batteries have not drawn much attention. Existing studies often require many parameters and cannot predict battery capacity and remaining life at the same time. Therefore, we propose a method that can jointly predict capacity and remaining life of retired lithium batteries using less parameters and data. Our method contains two steps: (1) capacity prediction using features extracted from parameters, (2) remaining useful life prediction using coefficient related to cycle index and former extracted features. Regression models are selected to complete two steps. In this article, basic definitions of parameters are given and datasets are introduced first. Then, battery health features containing mean value, skewness and kurtosis are extracted, as well as curve features with the concept of Frechet and Hausdorff Distance. Experiments are conducted for each type of feature by constructing Random Forest regressor, AdaBoost regressor and other 3 models on datasets, whose performances are measured by RMSE, MAE, MAPE and R2 score. The optimal model pair are selected as the prediction model.