Advancing Real-time Urban Flood Prediction through the Integration of Hydrodynamic Modeling and Spatiotemporal Machine Learning 8 views

Coastal cities are increasingly experiencing nuisance flooding (NF) due to the combined impacts of sea-level rise, intensifying rainfall, and urbanization. As its scale and frequency grow, NF imposes cumulative economic and societal costs. In coastal urban environments, flooding often occurs during compound events where elevated coastal water levels interact with precipitation. Consequently, coastal cities require tools to model these complex flooding processes within urban environments. Physics-based hydrodynamic models are widely used to analyze urban flooding because they can accurately simulate interactions among rainfall, tides, and drainage networks. However, these models are computationally expensive, limiting their application for real-time forecasting. Recent advances in machine learning (ML) techniques have enabled substantial reductions in computational cost while maintaining acceptable predictive accuracy. ML models offer the potential to serve as effective surrogates for physics-based hydrodynamic models. Building upon this context, this dissertation integrates physics-based modeling and ML approaches to enhance predictive performance and operational efficiency of urban flood prediction. Norfolk, Virginia, serves as the case study for this research. The first study employs a coupled one-dimensional and two-dimensional hydrodynamic model to quantify NF in Norfolk under present and future climate conditions. Simulations examine both tidal and compound nuisance flooding scenarios. Building upon the physics-based modeling output, the second study develops Long Short-Term Memory (LSTM) networks to predict flood depths for 40 flood-prone streets. The study evaluates different strategies for training ML models, including global models using all streets and clustered models grouping hydrologically similar streets. While LSTM models effectively capture temporal dependencies in flood dynamics, they do not represent spatial connections among locations. The third study enhances the forecasting framework by incorporating spatial relationships among streets using a Temporal Graph Convolutional Network (TGCN). This study proposes a weighted adjacency matrix that encodes spatial dependencies reflecting hydraulic relationships among streets using topographic features. Key findings from this research are that (1) projected sea-level rise and intensified rainfall expand the spatial extent of NF, with the affected area increasing threefold by 2100 compared to the 2020 baseline; (2) clustering streets based on hydrologic similarity for training LSTM-based flood models improves prediction accuracy compared to training using a single global model; and (3) incorporating spatial properties into a weighted TGCN improves prediction performance compared with baseline LSTM models that ignore spatial properties. Overall, this dissertation contributes to advancing the technical foundation for reliable flood prediction using physics-driven and ML modeling strategies to support flood resilience in coastal urban areas facing increasing flooding threats.

PHD (Doctor of Philosophy)

Real-time Flood Prediction; Machine Learning Surrogate Model

English

2026-04-22