A parallel LTCN-PHA network for remaining useful life prediction of lithium-ion batteries

Accurate prediction of the remaining useful life (RUL) of lithium-ion batteries (LIBs) is a critical component of battery health management systems. However, in practical applications, data scarcity poses a significant challenge due to the difficulty and cost associated with collecting comprehensive field data, often limiting the availability of battery capacity data for training. To address this issue, a parallel lightweight temporal convolutional network (LTCN)- parallel hybrid attention (PHA) model is proposed for accurate RUL prediction of LIBs. First, by refining the residual module of the original TCN, the proposed LTCN reduces both training parameters and computational time while enhancing prediction accuracy. Second, a parallel hybrid interactive channel attention and spatial attention enables the model to accurately capture the battery capacity decay trend and capacity regeneration phenomena. Finally, the parallel architecture ensures simultaneous and efficient extraction of degradation trends and regeneration features from limited battery capacity data. Comprehensive experiments were conducted on three varying lifetimes (short, medium, and long) battery datasets: on the NASA dataset, the model achieves a prediction error within 1 cycle in 5.2 s, with an RMSE below 0.02396 Ah; on the MIT dataset, the error remains within 1 cycle in 7.2 s, with an RMSE below 0.00249 Ah; and on the CALCE dataset, the error stays within 2 cycles in 15.7 s, with an RMSE below 0.01256 Ah. These results highlight the model's robustness and applicability across diverse battery aging scenarios.