Surrogate-assisted gray–green optimization for urban pluvial flood mitigation under rainfall variability and pipe siltation

Urban pluvial flooding is governed by the interactions among rainfall variability, drainage-system degradation, and mitigation strategies. However, repeated high-fidelity hydrodynamic simulations hinder efficient evaluation of diverse rainfall–infrastructure scenarios. This study develops a surrogate-assisted gray–green optimization framework integrating rainfall temporal patterns, pipe siltation effects, spatial flood prediction, and multi-objective planning. A Conditional Residual U-Net (CResU-Net) surrogate model was developed to reproduce spatial inundation fields from high-fidelity 1D 2D hydrodynamic simulations and benchmarked against multilayer perceptron (MLP), random forest (RF), and convolutional neural network (CNN) models. CResU-Net achieved superior cross-pattern predictive performance, with R 2 values of 0.961 and 0.957 under Mode II and Mode III rainfall scenarios, respectively, while preserving key spatial inundation characteristics. With an average inference time of approximately 9.0 s under the 100-year rainfall scenario, the surrogate model enabled efficient integration with the Borg Multi-Objective Evolutionary Algorithm (Borg MOEA) for gray–green infrastructure optimization. The results reveal that rainfall temporal structure reshapes flood responses by altering runoff accumulation and lateral connectivity. Early- and centrally peaked storms tend to generate larger inundation volumes, whereas late- and double-peaked storms promote spatial flood expansion. Pipe siltation intensifies flooding by reducing conveyance capacity and redistributing hydraulic loads, thereby amplifying node surcharge and overflow. The optimized Pareto fronts further identify a rainfall-dependent transition from conveyance-limited mitigation under light rainfall to storage–conveyance co-regulation under moderate rainfall and system-capacity exceedance under extreme rainfall. These findings provide a mechanism-based basis for adaptive gray–green flood mitigation under changing rainfall regimes and drainage degradation.