Battery Remaining Service Life Prediction Based on LSTM-WBO Method and Shapley Value

Predicting the remaining useful life (RUL) of lithium-ion batteries is central to battery health management. However, traditional physics-based models rely on precise electrochemical parameters and involve complex computations, while mainstream data-driven models suffer from the “black-box” limitation and struggle to adapt to the right-skewed errors commonly seen in RUL prediction, which restricts their practical application. Based on the NASA dataset, this paper proposes an interpretable prediction framework integrating “Long Short-Term Memory (LSTM), Weibull Distribution-Based Bayesian Optimization (WBO) and Shapley values”. Specifically, a two-stage strategy is adopted to select 8 key health factors. LSTM is used as the base prediction model, and WBO is employed to optimize hyperparameters for adapting to right-skewed errors. Additionally, Shapley values are utilized to analyze the contribution of features to prediction. Experiments show that the LSTM model significantly outperforms XGBoost, Random Forest, and the Adaptive Deep Neural Network reported in the literature. Shapley value analysis reveals that the charging current area is the key feature influencing RUL. This study balances prediction accuracy, error adaptability, addresses the “black-box” issue of data-driven models, and provides a reliable solution for battery maintenance in industrial scenarios.