Keyhole penetration is a key indicator of welding quality and process stability. Its accurate prediction is important for improving manufacturing efficiency and ensuring weld reliability. Although deep learning has advanced this task, most methods assume independently and identically distributed data, limiting performance across diverse welding scenarios. Moreover, traditional unsupervised domain adaptation mainly overlooks class-relevant information, restricting cross-domain generalization. To overcome these issues, we propose a Hierarchical Feature Alignment and Disentanglement (HFAD) framework for cross-domain keyhole penetration prediction. Specifically, in the first stage of HFAD, we jointly perform domain alignment to extract domain-invariant features and introduce a hybrid attention module to enhance their discriminability. In the second stage, we apply mutual information minimization to further extract class-relevant features, facilitating better adaptation to the target domain. When evaluated in three welding scenarios, the proposed method achieves state-of-the-art performance in cross-domain keyhole penetration prediction. Furthermore, ablation studies validate the effectiveness of each component in the HFAD framework.