An interpretable transformer model for battery health prediction based on a multitask learning framework and cross-attention mechanism

Accurate estimation of battery state of charge (SOC) and state of health (SOH) is essential for the safe control, effective scheduling, and reliable operation of electric vehicles and energy storage systems based on batteries. Although SOC and SOH are physically interdependent, they are still treated as separate estimation tasks in many existing studies, which limits prediction consistency and weakens the use of degradation information during charge-state estimation. To address this limitation, a Cross Attention-Based Multitask Transformer for Battery Health Prediction (CA-MT-BHP) is presented, meant for joint SOC and SOH valuation. The novelty of the suggested framework lies in three aspects: first, task-specific Transformer pathways are used to learn short-term electrical dynamics for SOC and long-term degradation behavior for SOH; second, a cross-attention mechanism is introduced to explicitly align SOC-specific and SOH-specific feature representations; and third, interpretability is incorporated through Local Interpretable Model-Agnostic Explanations to support transparent battery-state prediction. The model uses voltage, current, power, time, and battery temperature for SOC learning, while cumulative variables such as ampere-hour, watt-hour, chamber temperature, battery temperature, and time are used for SOH learning. The experimental results conducted on the Panasonic and NASA battery datasets show excellent predictive performance, with a testing mean absolute error (MAE) of 0.010% and testing root mean square error (RMSE) of 0.014% for SOC prediction, and a testing MAE of 0.008% and RMSE of 0.010% for SOH prediction, where R2 = 99.5% and R2 = 99.6%, respectively. The comparative and ablation analyses further demonstrate that explicit SOC–SOH feature alignment not only enhances robustness but also boosts the stability of the estimation. These outcomes validate the potential of the CA-MT-BHP framework to serve as an accurate, scalable, and interpretable methodology for more advanced battery management systems.