State-of-health estimation of lithium-ion batteries using a kernel support vector machine tuned by a new nonlinear gray wolf algorithm

• Published in: Journal of energy storage Vol. 102; p. 114052
• Main Authors: Liu, Shiyu, Fang, Lide, Zhao, Xiaoyu, Wang, Shutao, Hu, Chunhai, Gu, Fengshou, Ball, Andrew
• Format: Journal Article
• Language: English
• Published: Elsevier Ltd 15.11.2024
• Subjects: Electrochemical impedance spectroscopy; Kernel support vector regression; Lithium-ion battery; Machine learning; Nonlinear gray wolf algorithm; State of health; Lithium-ion battery; State of health; Nonlinear gray wolf algorithm; Machine learning; Electrochemical impedance spectroscopy; Kernel support vector regression
• ISSN: 2352-152X
• DOI: 10.1016/j.est.2024.114052

The computer-aided estimation of battery state of health (SOH) has been regarded as an active field of energy management because of the high demand for electric vehicles and consumer electronics. In this study, a new data-driven model is proposed for the capacity prediction and online monitoring of lithium-ion batteries, which is formulated based on a kernel support vector machine (KSVM) and a nonlinear Gray Wolf Optimization (NGWO) to capture the health information in electrochemical impedance spectroscopy (EIS) data. The amplitudes of EIS in the frequency range from 0.02 Hz to 20,000 Hz are taken as the input variables of KSVM model to predict the capacity at different cycles of battery charge-discharging. Moreover, GWO is improved through the proposed new inverse S-shaped exponential compound function convergence factor and position ratio-based dynamic weighting scheme to enhance its accuracy in optimizing KSVM parameters. The capacity prediction tasks of single battery (Case 1), different batteries at different temperatures (Case 2) and limited cyclic data (Case 3) are discussed in detail. Experimental results show that compared with other estimation methods, the NGWO-KSVM exhibit the lowest root mean square error (0.073 and 0.075 in Case 1, 0.434 and 0.263 in Case 2), the smallest mean absolute percentage error (0.052 and 0.055 in Case 1, 0.286 and 0.178 in Case 2), and the highest determination coefficient (0.936 and 0.956 in Case 1, and 0.981 and 0.993 in Case 2) for two different batteries in relatively short time. Also the NGWO-KSVM can more effectively utilize a fewer cycles of EIS data to improve capacity estimation performance in Case 3. It provides superior solution for the problem of low accuracy and poor robustness in battery capacity prediction, and has the potential for actual implementation in battery routine monitoring. •A new EIS-machine learning model is proposed to estimate battery SOH.•An inverse S-shaped exponential compound function is proposed to replace the linear convergence factor in GWO.•A position ratio-based dynamic weighting scheme is proposed to control the ranking system of GWO.•The proposed NGWO can improve the accuracy of optimizing KSVM parameters.•The effectiveness of the presented method is validated by three different conditions.