Statistical Post-Processing in Ensemble Learning-based State of Health Estimation for Lithium-Ion Batteries

Abstract

Using ensemble learning (EL) for battery state of health estimation has become a research hotspot. Because the performance of a single estimator can get boosted, which is applicable in the field of the battery especially when the amount of aging data is insufficient. Traditional EL is to aggregate base models through averaging, which will introduce errors from poor base models. To fully use the estimation results from base models, a statical post-processing method is proposed in this paper. The EL algorithm is initially constructed by combining random sampling and training multiple extreme learning machines. Then the post-processing is performed by fitting the kernel probability distribution of all sub-outputs and determining the most likely estimate, i.e., the statistical mode. As for comparison, the performance of other aggregations using average, weighted average, and mode from a normal distribution are investigated. Finally, the effectiveness of the proposed method is verified by conducting aging experiments on an NMC battery. The root-mean-squared error is as low as 0.2%, which is an approximate 80% improvement in accuracy over the traditional average-based method. The proposed method tackles the unstable estimation in learning with a small dataset, which is suitable for practical applications.