Capacity estimation of lithium-ion batteries based on optimized charging voltage section and virtual sample generation

Abstract

Accurate capacity estimation of a lithium-ion battery is crucial to predict its state of health and remaining useful life, determining the reliability and safety of the battery system directly. However, existing data-driven methods usually are based on full battery charging or discharging curves, which are hard to meet in practice because of range anxieties and arbitrary charging behaviors of users. Meanwhile, the lack of sufficient datasets for model training limits the performance of a data-driven model. In this paper, a capacity estimation method for lithium-ion batteries based on an optimized charging voltage section and virtual sample generation is proposed. In the method, characteristics of full and sectional capacity degradation are analyzed to evaluate the feasibility of capacity estimation using a section of charging voltage curve. The extraction of health indicator and optimization of battery voltage range are introduced. A multi-distribution global trend diffusion algorithm is used to generate virtual samples. Based on a long short-term memory neural network, battery capacity estimation model is constructed. Experimental verifications on NASA, CALCE, and Oxford battery datasets are carried out. Results show that the proposed method not only reduces the length of required voltage section for capacity estimation, but also improves the estimation accuracy in comparison to utilization of the full charging curve. Under the operating conditions of NASA B6, CALCE CS36 and Oxford C3, the root mean square error (RMSE) of battery capacity estimation is lower than 1.48%.