Battery State-of-Health estimation based on multiple charge and discharge features

• Published in: Energy (Oxford) 2023-01, Vol.263, p.125637-12, Article 125637
• Main Authors: Ospina Agudelo, Brian, Zamboni, Walter, Postiglione, Fabio, Monmasson, Eric
• Format: Article
• Language: English
• Subjects: Battery capacity indicators; Battery SoH estimation; Chemical Sciences; Electric power; Engineering Sciences; Features selection; Material chemistry; Mechanics; Multiple regression models; Variance inflation factor
• ISSN: 0360-5442; 1873-6785
• DOI: 10.1016/j.energy.2022.125637

In this work we propose and evaluate the performance of several linear multifeature models for battery State-of-Health estimation. The models combine high current incremental capacity and dynamic resistance features, which can be obtained during partial constant current charge and discharge, respectively. We construct the models by including fixed sets of features or by applying features selection procedures based on statistical criteria. The proposed models are fitted and evaluated with data from three publicly available battery datasets, including batteries cycled using driving, randomised and fast charging profiles. During the test process, we assess the estimation improvement introduced by each multifeature model by evaluating the reduction of the mean squared error in the State-of-Health estimation with respect to two reference single-feature models already used in recent literature. The collinearity is quantified through the variance inflation factor to indicate the prediction reliability of each model. As main result of this analysis, we propose a simple two-features model as the best compromise between estimation improvement with respect single feature models, and collinearity reduction. [Display omitted] •Several linear multifeature models for battery SoH estimation are proposed.•The models combine high current incremental capacity and resistance features.•The features only require partial charges and discharges.•The models performance is evaluated using data from three publicly available datasets.•The best performance was obtained with a two features model.