Lithium-Ion Battery State of Health Estimation Based on Model Interpretable Feature Extraction

This study proposes a lithium battery State of Health (SOH) estimation method that utilizes model interpretability feature extraction and the Equilibrium Optimizer (EO) algorithm to optimize Temporal convolutional neural networks (TCN), addressing issues of feature collinearity, noise interference, and the challenges of manual model parameter tuning. Initially, the battery’s incremental capacity (IC) curve is smoothed using Gaussian filtering, and the health features are extracted from the charging, discharging and IC curve to establish a TCN-based SOH estimation model. Subsequently, the SHAP interpretability method is employed to analyze the contribution of various features to the TCN model’s predictions, and the features were further screened based on these contributions; the EO algorithm is used to optimize the TCN model hyperparameters, enhancing the model’s prediction performance. Finally, this study builds an experimental platform for ageing tests to validate this method with experimental data and public datasets. The results show that SHAP analysis and the EO algorithm, based on the model’s real-time feedback mechanism, significantly improved the accuracy of feature selection and model prediction precision. The method proposed in this study achieves an RMSE of below 1.40% for the SOH prediction of six batteries, reducing it by 66.7% compared to the baseline model.