Interpretable machine learning prediction for li-ion battery's state of health based on electrochemical impedance spectroscopy and temporal features

Abstract

Accurately estimating lithium-ion battery's state of health (SOH) can effectively improve the safety and economics of energy systems, which is an unsolved challenge. Electrochemical impedance spectroscopy (EIS) is an information-rich battery data that can be measured in real time. Using interpretable machine learning tools can enhance the understanding and transparency of machine learning predictive models, seen as black boxes. This work integrates electrochemical impedance spectroscopy and temporal features derived from different electrochemical processes. Specifically, six machine learning models of varying complexity are employed to reveal the accuracy and reliability of battery health prediction. The mean absolute error (MAE) of the MLP model is 0.17, the mean squared error (MSE) is 0.32, the root mean squared error (RMSE) is 0.40, and the mean absolute percentage Error (MAPE) is 0.97 % with the used EIS features. After combining the features, the MAE of the XGBoost model is 0.05. The results demonstrate the effectiveness of our method and its interpretability.