Enhancing Battery's State of Health Estimation with Linear Models and Time- Energy Features

Estimating the State of Health (SoH) is crucial for the safety and reliability of batteries, with accurate SoH estimation heavily relying on the selection of appropriate features. This paper proposes using time & energy features to enhance prediction accuracy. Experiments were conducted using the NASA Li-Ion dataset, which includes various batteries with different charging and discharging profiles. Feature extraction considered power during constant current (CC) and constant voltage (CV) charging phases, time features in the charging-discharging process, and temperature. A feature selection method was employed, and both nonlinear and linear approaches were optimized using GridSearchCV for SoH estimation. Results indicate that linear models (Lasso, ElasticNet, Ridge, SVR with a linear kernel, linear neural network, GPR with a linear kernel) outperform nonlinear models (XGBoost, LightGBM, CatBoost, nonlinear neural network, and nonlinear SVR & GPR kernels) in accuracy and generalization. Among all tested methods, linear GPR, Lasso, and linear SVR provided the most accurate predictions, demonstrating that with proper feature extraction, simpler regression methods can accurately and quickly estimate SoH.