The emergence of high-voltage direct current integrated (HVDC) power systems introduces significant out-of-distribution (OOD) challenges to machine learning-based analysis of cascading failures (CFs). Existing algorithms often suffer from poor interpretability and generalization in handling diverse CF propagation modes. To address these issues, this paper proposes a novel Causal Graph Attention (CGAT) model that captures critical information from CF events to accurately predict propagation paths and impacts. Specifically, by employing a graph attention mechanism, the model extracts both causal and trivial features as distinctive attention dual subgraphs, which aids in differentiating shortcuts in causal variables and improving prediction accuracy. These subgraphs also provide visual representations that highlight the key elements influencing CF propagation. By incorporating interventions, the CGAT model enhance prediction capabilities across a variety of scenarios that include HVDC control strategies and renewable energy sources (RES). Extensive experiments on augmented IEEE 30-bus and 2737-sop systems demonstrate the effectiveness and superiority of the proposed CGAT model.