| With the development of the economy and society,countries around the world are facing problems such as energy shortages and climate change.Exploring the path to sustainable development has become a consensus.Against the background of the "dual carbon" target,accelerating the development of new energy and realizing electrification has become the main trend of China’s energy clean and low-carbon transformation.Lithium-ion batteries have advantages such as high efficiency,light weight,high energy density,and environmental friendliness,and are widely used in electric vehicles,3C products,energy storage,and other fields.Therefore,they are the key technology for establishing and applying new power systems.However,they also face the problem of irreversible electrochemical performance degradation,which is nonlinear and timevarying,posing a challenge to the safety and stability of the system.Therefore,developing efficient and reliable battery management systems(BMS)to accurately estimate the battery’s state of health(SOH)has important research value.Among them,the data-driven machine learning method,compared with the model-based method,has the advantages of not relying on expert knowledge,low computational load,and strong adaptability,and is currently a hot topic in battery SOH estimation research.Mining healthy factors(HIs)with high robustness and ability to describe battery degradation states is the key to achieving accurate estimation using machine learning methods.To improve the robustness of HIs,this paper starts from the constant voltage(CV)charging phases of lithium-ion batteries,which is more stable and easier to obtain.Statistical features are used to describe the capacity curve of CV charging phases,and Pearson correlation coefficient(PCC)evaluation finds that only the statistical features of the capacity curve extracted from the beginning 4 minutes of the CV phases can accurately describe the aging state of the battery.A powerful and efficient Tree Parzen Estimator(TPE)optimization algorithm is proposed to tune hyperparameters of XGBoost tree algorithm and SVR to find the implicit relationship between HIs and SOH.Two standard functions are used to prove the feasibility and accuracy of using short-term capacity HIs from constant voltage phases for battery SOH estimation with TPE-XGBoost method.Functional Analysis of Variance(FAOV)algorithm and SHapley Additive explanation(SHAP)algorithm are used to analyze the key hyperparameters and HIs that affect the machine learning estimation performance,to solve the "black box" model problem of complex and multidimensional machine learning algorithm estimation of battery SOH.Furthermore,according to the literature P2D model simulation results,the battery resting voltage relaxation process is the external manifestation of the three internal polarization modes of the battery.Therefore,the voltage relaxation process contains direct information on ion and electron transport within the battery and can be easily obtained.A least squares method combined with a second-order RC equivalent circuit model is proposed to fit the voltage relaxation process,obtaining impedance HIs related to polarization,which have the clear electrochemical meaning.The TPE-XGBoost method is used to obtain the relationship between impedance HIs and battery SOH.The evaluation finds that compared with the TPE-XGBoost aging battery estimation model based on CV short-term capacity HIs,the average root mean square error(RMSE)of the TPE-XGBoost model based on impedance HIs is reduced by 45%,and the RMSE is smaller than 0.0028,with a determination coefficient(R2)greater than 0.999,indicating a significant improvement in SOH estimation accuracy.Combined with the SHAP algorithm,the key features of the XGBoost model based on impedance HIs are found to be sourced from the voltage relaxation phases after discharge,which has a more significant impact on battery aging than these based on voltage relaxation process after CV charging phases. |
PDF Full Text Request