Lithium-ion battery state-of-health estimation based on health feature selection and IGWO-Transformer-GRU model

Abstract

This study aims to address the accuracy and robustness limitations of conventional single-model approaches for lithium-ion battery state-of-health (SOH) estimation. This paper proposes an enhanced improved Grey Wolf optimization (IGWO)-Transformer-GRU-based estimation framework. Health factors (HFs) of lithium-ion battery cells are extracted from a set of signals, such as charge/discharge current, terminal voltage, body temperature during the charging stage and the incremental capacity curve. Kernel principal component analysis and Pearson correlation analysis are used to reduce the dimensionality of these HFs. Moreover, to enhance the global search and convergence capabilities of the SOH estimator, Tent chaotic mapping and Lévy flight are introduced to improve the estimation processes, and a Transformer-GRU model is developed to form a remaining useful life prediction framework. Ablation studies and comparisons are provided based on NASA and CALCE battery datasets to validate the proposed method. The proposed IGWO-Transformer-GRU method is compared with some benchmark algorithms, such as Transformer, Transformer-GRU, IGWO-GRU and GWO-Transformer-GRU models on the NASA and CALCE battery datasets. According to the comparison results of eight groups of tested batteries, the proposed method achieves average mean absolute error (MAE) and root mean squared error (RMSE) values of 0.00675 and 0.00846, respectively. The proposed method outperforms the Transformer, Transformer-GRU, IGWO-GRU and GWO-Transformer-GRU models, reducing the average MAE by 60.06%, 49.20%, 31.12% and 28.76%, and the RMSE by 57.76%, 46.82%, 30.64% and 28.13%, respectively. These quantitative results demonstrate that the proposed IGWO-Transformer-GRU model provides superior SOH estimation accuracy and robustness. This study proposes a novel hybrid framework for the SOH estimation of lithium-ion batteries, which enhances global search capability and convergence velocity by introducing a series of improved feature selection mechanisms. Furthermore, this paper builds up a Transformer-GRU hybrid scheme to capture both long-term temporal dependencies and local degradation patterns of batteries. This scheme presents a robust solution for battery management systems, significantly advancing the estimation accuracy and reliability.