Prognostics of battery capacity based on charging data and data-driven methods for on-road vehicles

Abstract

The large-scale application of lithium-ion batteries makes it urgent to accurately predict their capacity degradation so as to achieve timely maintenance and second-life utilization. For on-road electric vehicles (EVs), due to limitation of battery management system in measurement and computing power, it is still a tricky challenge to accurately predict the capacity of battery pack. To this end, a battery capacity prognostic method based on charging data and data-driven algorithms is proposed in this paper. First, battery capacity is calculated based on a variant of Ampere integral formula, and statistical values of the capacity during a month are regarded as labeled capacity to reduce errors. Then, statistical characteristics of battery charging data are extracted, and correlation analysis and feature selection are conducted to determine optimal feature sets. Moreover, a sequence-to-sequence (Seq2Seq) model is employed to predict future capacity trajectory, and two residual models based on Gaussian process regression (GPR) are proposed to compensate the prediction error caused by local capacity change. Finally, the data of 20 EVs operating about 29 months are used to verify the proposed methods. By using the first 3 months data as input, the remaining capacity sequence can be accurately predicted with error lower than 1.6%.