State of health estimation of lithium-ion batteries using variational-diffusion generative model and attention-enhanced hybrid temporal network

State of health (SOH) of lithium-ion batteries is essential for maintaining the safety and reliability of energy storage systems. However, conventional deep learning-based prediction methods heavily depend on high-quality data and often struggle to capture the complex temporal degradation behavior of batteries. To overcome these challenges, this study proposes an SOH estimation framework based on a variational-diffusion generative model (VDGM) and an attention-enhanced hybrid temporal network (AEHTN). First, a collaborative generative model, termed VDGM, is introduced. Specifically, a variational autoencoder (VAE) provides global prior information for the denoising diffusion process and supplies prior-informed initialization for the reverse denoising process. This design enables VDGM to capture both coarse-grained degradation trends and fine-grained feature details simultaneously, thereby generating high-quality degradation data. Next, an AEHTN is developed by integrating the bidirectional causal convolution architecture with an exponential gating mechanism and a dynamic feature weighting strategy. This design allows the AEHTN to adaptively capture both short-term and long-term critical degradation features. Finally, high-quality synthetic samples are selected according to feature correlation and integrated into the training process, which enriches data diversity and significantly improves SOH estimation accuracy. Experimental results on the CALCE and NASA datasets demonstrate that the proposed framework outperforms state-of-the-art approaches in both data generation quality and SOH estimation performance.