An attention-enhanced multi-scale feature fusion framework for accurate joint prediction of battery SOH and RUL

Precise prediction of state-of-health (SOH) and remaining useful life (RUL) for lithium-ion batteries (LIBs) is pivotal to optimizing the battery management system (BMS) and advancing sustainable energy systems, yet remains challenging due to local steep drops and capacity regeneration phenomenon. To solve the limitations, this work presents a hybrid deep learning framework that integrates the temporal convolutional network (TCN), bidirectional gated recurrent unit (BiGRU), and attention mechanism (AM). The TCN extracts multi-scale degradation features and long-term trends of LIB capacity degradation, while the BiGRU accurately captures the local fluctuations and short-term mutations in the degradation process, and the AM prioritizes critical phases through adaptive weight allocation, thereby mitigating interference from redundant data. Evaluated on CALCE-CS2 dataset, the proposed TCN-BiGRU-AM model achieved an average R² exceeding 0.98, confirming its remarkable performance. Notably, our model yields a 71.5% reduction in MAE for joint prediction relative to the single BiGRU model. Moreover, our model maintains accuracies above 96% on both the NASA-18650 and CALCE-CX2 datasets, substantiating its strong generalization capability across diverse battery types. This work presents a scalable solution for accurate SOH and RUL prediction, facilitating advancements in energy storage and electric mobility applications.