期刊
ETRANSPORTATION
卷 15, 期 -, 页码 -出版社
ELSEVIER
DOI: 10.1016/j.etran.2022.100214
关键词
Lithium -ion battery; SOH; RUL; Differentiation; Machine learning; Cloud
This paper utilizes feature-based machine learning techniques to predict the capacity fade of large format LiFePO4 batteries in EV applications. Through noise reduction, feature engineering, and stacking ensemble learning, accurate estimations and predictions of battery capacity and remaining useful lifetime are achieved based on field data collection.
Predicting the evolution of multiphysics battery systems face severe challenges, including various aging mechanisms, cell-to-cell variation and dynamic operating conditions. Despite significant progress, solving real-life battery problems with noisy and missing data and high-dimensional parameter space are either difficult or impossible. In this paper, we design and evaluate feature-based machine learning techniques for estimating the capacity of large format LiFePO4 batteries in EV applications and hence predicting the trajectory of capacity fade based on the estimations. To probe the feature space, we generate a comprehensive dataset consisting of 420 cells and 9 battery packs (178 cells in-series for each one) with more than 10,000 validation data derived from the cloud platform. A two-step noise reduction method is applied to de-noise the scattered field data (30s sampling interval). Totally, 39 domain features are engineered using the reconstructed segments of battery charging data based on the differential methods (increment capacity and differential voltage), which steer the learning process towards accurate and physically consistent predictions by leveraging the stacking ensemble learning. The stacking, comprised of four base learning models referred to as level-1 predictors and a meta-learner referred to as level-2 predictor applied to combine predictions of base learners with probability distributions using an extended set of meta-level features offers exciting opportunities for better accuracy and improved generalization. Our best models achieve 0.28% mean absolute percentage error (MAPE) and 0.55% root mean squared percent error (RMSPE) for battery capacity estimation. Further, 1.22% average percentage error is achieved in the prediction of remaining useful lifetime (RUL) under different conditions of driving distance (km) and service time (day) by building capacity fade trajectory based on a Bayesian regression. This work highlights the promise of machine learning modelling using domain-specific features for accurate estimation and prediction of real-life battery systems based on field data collection.
作者
我是这篇论文的作者
点击您的名字以认领此论文并将其添加到您的个人资料中。
推荐
暂无数据