Explainable Hybrid Deep Learning for Lithium-Ion Battery Health Assessment and Lifetime Prediction

Abstract

The increasing deployment of lithium-ion batteries in electric vehicles, renewable energy storage systems, and portable electronics necessitates intelligent Battery Management Systems (BMS) for reliable assessment of battery degradation. State of Health (SoH) and Remaining Useful Life (RUL) are critical prognostic indicators; however, their estimation remains challenging due to nonlinear degradation characteristics and the limited interpretability of deep learning models. This study develops an explainable deep learning framework based on a CNN–LSTM architecture for joint SoH and RUL prediction. Battery charge and discharge measurements are temporally aligned, normalized, and transformed into multivariate sequences comprising voltage, current, and temperature. The CNN component extracts localized temporal representations, while the LSTM component captures long-term degradation dependencies. Standalone LSTM and GRU models are employed as baseline architectures for comparative evaluation. Integrated Gradients is subsequently applied as a post-hoc Explainable Artificial Intelligence (XAI) method to quantify the contribution of input features to model predictions. Experimental results demonstrate that CNN–LSTM achieves superior predictive performance, with an MAE of 0.571, RMSE of 0.811, MAPE of 6%, and R2 of 0.94. The XAI analysis indicates that discharge current at later timesteps contributes substantially to the model predictions, providing interpretable evidence of degradation-related patterns. The proposed framework consequently integrates predictive accuracy and post-hoc interpretability to support reliable data-driven battery prognostics for intelligent BMS applications.