Improving Real-Time Street-Scale Flood Forecasting in Coastal-Urban Environments using Surrogate Modeling, Transfer Learning, and Physics-Informed Machine Learning 7 views

Coastal-urban environments are facing an escalating threat of nuisance flooding driven by heavy rainfall, sea-level rise, urbanization, and inadequate stormwater drainage infrastructure. Although short-duration nuisance flooding is less destructive, it frequently disrupts transportation systems within the cities. This highlights the need for real-time street-scale flood forecasting to make decisions on flood warning and emergency response for transportation infrastructure. Physics-Based Models (PBMs) are commonly used for flood predictions, but they require high computational runtimes that are not well-suited for real-time decision-making. Machine Learning (ML) models, on the other hand, offer faster runtimes, but they are less trusted by decision-makers who prefer PBMs. To address the challenge of computationally intensive PBMs and less trusted ML models, surrogate models trained from high-fidelity PBMs are employed as a practical solution to enable real-time flood prediction. Hence, the first study explores the potential of a seq2seq (sequence to sequence) Long Short-Term Memory (LSTM) surrogate model for short-term (4-hr) and long-term (8-hr) street-scale flood forecasting and compares its performance with a traditional LSTM surrogate model. While the surrogate models provide accurate results, it is still questionable if the trained ML model can be transferred to new locations without requiring retraining (or with only minor or full retraining). To address this, the second study explores different Transfer Learning (TL) approaches to improve the utility of LSTM models for street-scale flood forecasting by transferring knowledge from data-rich streets to data-poor streets. Finally, while the surrogate models achieve accuracy and transferability in flood forecasting, they are still unconstrained by the underlying physical processes behind flood routing. Hence, the third and final study explores Physics-Informed Neural Networks (PINNs) by integrating physics laws along with the ML surrogate models for street-scale flood forecasting. Specifically, we explored flexible PINNs that incorporate physics-based prior information into the loss function of the neural networks, without requiring the continuous calculation of derivatives. For all these studies, the city of Norfolk, Virginia was used as the case study due to its increased risk of nuisance flooding from climate change and sea level rise. Key findings from this research are that (1) the seq2seq LSTM surrogate model showed lower error and higher accuracy than the LSTM surrogate model across most lead times, particularly for long-term forecasting; (2) TL improves forecasting accuracy when target data are limited (less than 18 storm events) by transferring knowledge from a model pre-trained on the source dataset; and (3) incorporating a mass balance equation into the LSTM surrogate model as a physics-based loss improved flood-depth forecasting accuracy by 13.9%. Taken together, this dissertation presents novel methods for integrating PBMs and ML to advance real-time, street-scale flood forecasting, providing reliable inundation information that supports decision-makers and enhances the resilience of transportation infrastructure under increasing flood risk.

PHD (Doctor of Philosophy)

Flood Forecasting; Urban Hydrology; Street Flooding; Machine Learning; Surrogate Modeling; Transfer Learning; Physics-Informed Machine Learning; Real-Time Prediction

English

2026-04-22