Lithium-ion battery state of health estimation using meta-heuristic optimization and Gaussian process regression

Abstract

Wrapper methods are widely employed in feature selection for status prediction of lithium-ion batteries and Gaussian process regression (GPR) is often adopted for state-of-health (SOH) estimation. However, the number and the source of the features are not considered as constraints in the wrapper methods, most existing optimization algorithms for finding optimal hyperparameters of GPR easily get trapped into the local optimum. In this work, a newly developed meta-heuristic optimization algorithm, Hunger Game Search (HGS) is utilized, because of its robustness and competitive performance to find the solution of both constrained and unconstrained problems. Also to further improve HGS the chaos mechanism is embedded to form Chaos-HGS (CHGS). It is utilized to find global optimal hyperparameters in GPR and its binary variant (CBHGS) is adopted to solve feature selection problems. Effects of chaos maps and constraints upon SOH prediction and hyperparameters optimization are investigated. To gain a deeper understanding of wrapper method results, intrinsic characteristics of features are mined through Gaussian Process Latent variables model (GPLVM). Relationships between features and the capacity of the battery are examined through clustering analysis. The effectiveness of the proposed algorithms and data mining methods is verified on the lithium-ion battery dataset of the NASA Prognostics Center of Excellence. The results show that the features found by CBHGS working with GPR can provide SOH prediction with higher accuracy and lower cost.