A Comparative Study of Deep Time-Series Models for Lithium-Ion Battery State-of-Health Prediction

Model rankings in battery SOH prediction can change with the input representation and battery split. This study therefore evaluated five deep time-series models—Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), PatchTST, TimesNet, and ModernTCN—on the NASA battery aging dataset under chronological and cross-battery protocols using the same input pipeline. The chronological protocol predicts later cycles of B0005, B0006, and B0007, whereas the cross-battery protocol trains on these three batteries and evaluates on the unseen B0018 battery. All models use the same 20-cycle input window and 14 cycle-level features derived from voltage, current, temperature, discharge duration, and historical capacity. The prediction target is the one-step-ahead SOH, defined as \(C_t/C_0\). Each configuration was trained with five random seeds (42–46), and the results are reported as the mean ± sample standard deviation. Under chronological evaluation, GRU achieved the lowest mean MAE, \(0.02513 \pm 0.00708\), while PatchTST obtained the lowest mean RMSE, \(0.03429 \pm 0.00851\), and the highest mean \(R^2\), \(0.73132 \pm 0.13911\). On the unseen B0018 battery, TimesNet performed best, achieving an MAE of \(0.01726 \pm 0.00394\), an RMSE of \(0.02213 \pm 0.00438\), and an \(R^2\) of \(0.88631 \pm 0.04139\); ModernTCN produced closely comparable results. The best model depended on the split: GRU achieved the lowest chronological MAE, whereas TimesNet achieved the best mean performance on the unseen B0018 battery. Removing historical capacity improved some architectures but degraded others, and the direction of the change also depended on the evaluation protocol. Because the cross-battery evaluation uses one held-out battery, these findings should be interpreted as evidence of protocol sensitivity rather than as a universal ranking of the models.