A GWO-BiLSTM Model-Based Method for Lithium-Ion Battery SOH Prediction Using Charging Phase Data

To address the challenges of limited feature extraction and low prediction accuracy in the estimation of lithium-ion battery state of health (SOH), this study presents a novel SOH estimation approach leveraging charging phase data and a grey wolf optimizer (GWO)-enhanced bidirectional long short-term memory (BiLSTM) network. The method begins by extracting five categories of health features (HFs) from the charging data, capturing key aspects of the battery’s operational behavior. To reduce dimensionality and identify critical indicators, kernel principal component analysis (KPCA) is applied, isolating the most informative features for SOH prediction. Subsequently, a GWO-BiLSTM framework is constructed, wherein the GWO algorithm serves as a global optimizer, dynamically searching for the optimal combination of BiLSTM hyperparameters—such as learning rate, regularization strength, and number of training epochs—by emulating the hunting strategy of grey wolves. This optimization process aims to minimize the SOH prediction error, enhance model convergence, and improve overall estimation accuracy. With the optimal hyperparameters determined, the BiLSTM model is trained to capture the complex, nonlinear mapping between battery health characteristics and SOH, facilitating precise SOH estimation. The proposed approach is validated using the NASA battery aging dataset, demonstrating the model’s ability to deliver highly accurate SOH predictions with a root mean square error (RMSE) consistently maintained below 1%, confirming its robustness and reliability.