Enhancing degradation trend prediction of lithium-ion battery capacity in complex aging scenarios: A Bayesian-optimized hybrid architecture combining local and global feature learning

Accurate prediction of lithium-ion battery capacity degradation under complex aging conditions is essential for reliable health monitoring in energy storage systems. Existing prediction methods exhibit limited capability in resolving the multi-scale measurement challenge for simultaneously capturing long-term degradation trends and short-term capacity fluctuations. This study develops a Bayesian-optimized Transformer-CNN hybrid architecture (BOTC) that innovatively integrates multi-head self-attention mechanisms for global trend measurement and adaptive 1D convolutional kernels for local anomaly quantification, incorporated with Gaussian process surrogate modeling for Bayesian hyperparameter optimization and dual sliding-window sampling strategy. Furthermore, systematic ablation experiments integrated with Shapley value theory establish a quantifiable contribution model for each module in the hybrid architecture. Rigorous experimental validation on datasets of two distinct chemical systems was performed to prove the validity of the method. Results demonstrate that the proposed feature fusion architecture achieves significant improvements in both accuracy and robustness compared to advanced baseline models. Contribution quantification analysis reveals complementary mechanisms among Transformer, CNN, and Bayesian optimization in long-term trend modeling, local fluctuation detection, and model robustness enhancement. The proposed framework provides a high-precision and interpretable solution for battery health management, while its modular design enables task-specific customization, offering broad engineering applicability in real-world energy storage systems.