Data-Driven SOH Early Prediction of Lithium-ion Batteries

By lowering the number of cycles required in battery testing, time and expenses can be considerably reduced. High-dimensional datasets of lithium-ion batteries, containing numerous features, may lead to overfitting, increased computational complexity, and reduced model interpretability. Removing irrelevant or trivial features can enhance accuracy, generalization, and robustness of the model. Additionally, training times and computational costs are significantly reduced. In this study, a feature-based data-driven technique is designed to estimate the state of health of lithium-ion batteries early. This work focuses on the early prediction of the state of health (SOH) of Li-ion batteries using geometrical and statistical features from the first five diagnostic cycles. Elastic net, support vector regression (SVR), and linear regression are employed, with principal component analysis (PCA) applied within a machine learning pipeline, to extract the optimal predictive model. The INR21700-M50T battery cells, featuring a graphite/silicon anode and nickel-manganese-cobalt cathode, are used to verify the models. The tests spanned 23 months and utilized the Urban Dynamometer Driving Schedule (UDDS) discharge driving profile, along with a constant current-constant voltage (CC-CV) charging protocol at varying rates from C/4 to 3C. Measurements from capacity tests performed during reference performance tests or diagnostic cycles are used to construct input features for a machine learning model. The results demonstrate that the proposed algorithm offers several advantages: lower computational cost, higher performance, more excellent parameter identification stability, and more straightforward parameter settings. The last and highest MAE of 0.003 and 0.096 ampere-hours are recorded, respectively.