Climate change and human activities are intensifying the hydrologic cycle and increasing extreme events, challenging accurate prediction. This study builds on the Transformer architecture by introducing a sliding time window and runoff classification mechanism, enabling high-precision long-term runoff forecasting and significantly improving the simulation of extreme floods. However, the generalization ability of data-driven models remains limited in non-stationary environments. To address this issue, we further propose a hybrid framework that couples the process-based GBHM with the enhanced Transformer via bias correction. This fusion leverages the strengths of both models: the process-based model explicitly captures topographic heterogeneity, the spatial distribution of meteorological forcings, and their temporal variability, while the data-driven model excels at uncovering latent relationships among hydrological variables. The results demonstrate that the coupled model significantly outperforms traditional approaches in peak-flow prediction and exhibits superior robustness and generalizability under changing environmental conditions.

Urban underground spaces are rapidly expanding, but their low elevation, limited drainage capacity, and strong enclosure make them highly vulnerable to pluvial flooding. To elucidate how inundation dynamics in 3D underground spaces under extreme rainfall translate into actionable risk indicators (e.g., depth thresholds and arrival time), we propose and cross-validate a rainfall-informed capacity–depth–damage (C–D–D) curve method and a physics-based computational fluid dynamics (CFD) inundation model. The first approach is a rainfall-informed C–D–D curves method that rapidly maps net inflow to depth evolution and warning indicators (e.g., threshold depth and arrival time). The second approach is a 3D-geometry-resolved CFD inundation model that simulates spatially distributed depths/flows under prescribed inflow and drainage/outlet conditions, providing high-fidelity validation and hazard maps. A GeoSLAM handheld 3D Laser Scanning system was used to reconstruct as-built, modeling-ready 3D geometry of the underground space, addressing the common limitation of idealized layouts in prior evacuation-time assessments and enabling geometry-specific inundation and warning-threshold predictions. Using an underground parking garage in Tongzhou District, Beijing as a case study, we evaluated flood dynamics and risks under rainfall scenarios with annual exceedance probabilities of 1%, 2%, and 5%. Results show that stronger rainfall significantly advances critical water-depth thresholds and compresses evacuation windows; for example, under P = 1%, the 0.2 m alert occurs 1.5 h earlier than under P = 2% and 5.3 h earlier than under P = 5%. The two methods exhibit strong consistency in threshold timing (typically within 0–1 h), while CFD resolves spatial heterogeneity and identifies medium-to-high risk zones earlier in the intrusion stage. This integrated framework supports rapid early warning, evacuation-window assessment, entrance protection, and drainage-capacity design. Novelty lies in (i) integrating handheld 3D Laser Scanning with a “curve-first, CFD-refine” dual-model workflow; (ii) cross-validating fast C–D–D-based warning thresholds against geometry-resolved CFD dynamics; and (iii) delivering actionable time-to-threshold warnings and spatial risk maps for emergency planning.