A physically constrained proxy framework considering a process-aware gating mechanism for urban flood simulation.

Area of Science:

• Environmental Fluid Dynamics
• Computational Hydrology
• Machine Learning for Environmental Science

Background:

• Urban pluvial flooding results from complex interactions between rainfall and drainage systems.
• Existing models struggle with hydrological memory and cumulative flood effects.
• Accurate flood forecasting is crucial for urban risk management.

Purpose of the Study:

• To develop a process-aware and physically-constrained surrogate model for 2D urban surface inundation.
• To improve the prediction of water depth and flow velocity by incorporating hydrological memory.
• To provide a robust framework for real-time urban flood forecasting.

Main Methods:

• Integration of 1D drainage overflows (SWMM) as dynamic forcing for a 2D shallow water equation (SWE) surrogate model.
• Development of a PG-CNN-LNN model using hydrological process indicators (cumulative overflow, rainfall intensity) to capture memory effects.
• Implementation of Closed-form Continuous-time (CfC) dynamics and a negative water depth penalty for stability and physical consistency.

Main Results:

• The PG-CNN-LNN model achieved R² values > 0.98 for water depth and > 0.92 for flow velocity.
• Average absolute error for water depth and flow velocity was reduced by at least 50.0% and 28.6% compared to LSTM.
• Physical constraints significantly reduced negative water depth violations from 34.8% to 0.49%.

Conclusions:

• The proposed framework offers a significant advancement in urban flood inundation modeling.
• The model's accuracy and robustness are validated through superior performance against benchmark models.
• This physically traceable approach enhances real-time flood forecasting capabilities in complex urban environments.