A Stacking Ensemble Model Method For Rapid State-of-health Estimation Of Lithium-ion Battery

Accurate state of health(SOH)estimation is a fundamental prerequisite for the safe and reliable operation of lithium-ion batteries in scenarios such as energy storage and electric vehicles,and key technology to facilitate the commercialization of energy storage and electric vehicles.While guaranteeing accuracy,it is also quite important to improve the flexibility and rapidity of SOH estimation,which plays an active role in applications such as battery recycling and consistency assessment.In addition,the robustness of the SOH estimation method is particularly important due to the presence of measurement errors and environmental noise.Therefore,the key challenge is to improve flexibility and rapidity while guaranteeing accurate SOH estimation and to improve the robustness of the method.This paper presents an in-depth study of SOH estimation techniques for lithium-ion batteries.In order to explore the principles and advantages and disadvantages of different SOH estimation methods,as well as to analyze the problems in the current research,this paper firstly investigates the current SOH estimation methods for lithium-ion batteries.Data-driven machine learning methods can combine the advantages of high accuracy and low computational requirements,but it is difficult to guarantee the flexibility and rapidity of SOH estimation,and the accuracy and robustness could be improved.To improve the flexibility and rapidity of the machine learning method,the variation pattern of the voltage curve of the constant current charging stage of the battery under different aging conditions is first analyzed.Eight different short-term voltage profiles with a time window of only 300 seconds are divided based on the voltage curve.In order to accurately characterize the short-term voltage profiles,a series of time-domain statistical features are proposed,and the rationality of different short-term voltage profiles and different time-domain statistical features is verified by the Kendall correlation coefficient.By dividing multiple short-term voltage profiles and selecting suitable and short time window,the proposed method can performe SOH estimation at multiple charging process stages with a short time duration required for feature extraction,improving the flexibility and rapidity of SOH estimation.To address the situation of low accuracy and robustness of models in machine learning methods,this paper proposes a stacking ensemble model,which contains both robust models,such as light gradient boosting machine(light GBM),extreme gradient boosting(XGBoost),random forest(RF)and support vector regression(SVR),as well as a model with high accuracy,such as Gaussian process regression(GPR).The accuracy and robustness of the method are further enhanced by the accurate and efficient search for the optimal hyperparameter values for each model through Bayesian optimization.Experimental validation of the stacking ensemble model is carried out with two different datasets,showing that the method has the highest accuracy compared to lightGBM,XGBoost,RF,SVR and GPR models,with an average root mean square error(RMSE)of less than 1.5%.The method has also achieved the most accurate or more precise estimates for eight different short-term voltage profiles.In comparison,the method has the highest accuracy under different standard deviation Gaussian white noise interference.In addition,the method has the lowest average RMSE when compared to other models under the fast charging conditions and also has the best noise immunity.