Two-stream Temporal Fusion Network for Li-ion Batteries SOH Estimation

Accurately estimating the State of Health (SOH) of lithium-ion batteries has become a central research focus in the field of energy storage and management. However, most existing data-driven approaches primarily focus on intra-cycle features within a single charge-discharge cycle, often overlooking the intercycle correlations that capture long-term degradation patterns. To address this limitation, we propose a Two-Stream Temporal Fusion Network (TSTFN) that simultaneously captures finegrained intra-cycle sampling features and long-range inter-cycle temporal variations. Specifically, the Intra-Cycle Feature Encoder (ICFE) is designed to model intra-cycle dynamics and the InterCycle Trend Encoder (ICTE) to extract inter-cycle correlations, enabling a more comprehensive representation of battery degradation. Furthermore, we design a Degradation-Aware Fusion (DAF) module to adaptively integrate the complementary feature representations into a unified latent space. To enhance interpretability, we segment the battery degradation process into four distinct stages, assigning varying levels of importance to each stage based on its degradation maturity. Extensive experiments on the NASA battery datasets demonstrate that the proposed TSTFN achieves superior performance compared to state-of-the-art methods, highlighting its effectiveness in capturing both shortterm fluctuations and long-term degradation trends for accurate SOH prediction.