Accurate health status prediction of lithium-ion batteries is crucial for enhancing the safety and reliability of battery management systems. However, due to the complexity of the battery degradation process and the influence of different working conditions, existing approaches are inadequate in feature extraction and cross-domain prediction. To address these challenges, this paper proposes a health status prediction approach based on multi-scale spatio-temporal features and hierarchical domain alignment. The proposed approach utilizes residual multi-scale convolution and variational attention-based bidirectional long and short-term memory networks to jointly model local fine-grained variations and global evolutionary trends during battery degradation. Furthermore, a hierarchical domain alignment mechanism is employed to mitigate data distribution shifts across different working conditions by layer-wise discrepancy minimization and adversarial feature regularization in a complementary manner. Finally, uncertainty estimation based on Bayesian inference is incorporated to provide a credibility measure for the capacity prediction results. Experimental results demonstrate the accuracy and robustness of the proposed approach for capacity prediction under diverse transfer tasks and show strong cross-condition generalization and application potential.