• Introduces a DRO-based active search strategy that builds a robust probabilistic ambiguity set to identify critical security boundary points, balancing uncertainty misestimation and conservatism.
• Applies a Transformer-based deep learning model with self-attention to capture global nonlinear dependencies and reconstruct the full high-dimensional security region efficiently.
• Demonstrates accurate characterization of security regions on IEEE test systems at low computational cost, with strong robustness to renewable-load uncertainty.
• Provides a new data-driven paradigm for power system security assessment and decision support under high renewable uncertainty.