Real-Time Data-Driven Urban Flood Forecasting

Flood forecasting is a critical component of mitigating risks to human life, infrastructure, and ecosystems, particularly as climate change increases the frequency and severity of extreme weather events. Especially urban areas are vulnerable due to high sealing rates, dense infrastructure, and concentrated populations. Traditional forecasting approaches rely on physically based hydrodynamic models that provide accurate information on the location, timing, and extent of flood events. While these models represent the state of the art, they are computationally too demanding for real-time forecasting and operational decision support. An alternative is offered by data-driven models, which learn patterns and relationships directly from the data. By approximating the underlying input–output dynamics without solving the governing physical equations explicitly, data-driven approaches achieve substantially reduced computation times, making them suitable for real-time flood forecasting.
This dissertation develops and advances data-driven forecast systems tailored for urban environments, addressing three major flood types: (1) sewer surcharge events, (2) fluvial floods, and (3) pluvial floods. The research follows a structured progression, advancing from localized short-term predictions to large-scale dynamic forecasts, and finally to domain-independent transferable, physically informed frameworks.
For sewer surcharge events, an Ensemble Forecast System based on nonlinear autoregressive networks with exogenous inputs was developed. The Ensemble Forecast System predicts the onset and duration of surcharge events in manholes while explicitly quantifying structural model- and parameter uncertainty through ensemble modeling. With observed events consistently captured within uncertainty bounds, the system provides a reliable framework for early warnings and can support model-based or real-time control in modern sewer networks.
For fluvial floods, a Feature-Informed Forecast System was introduced, framing flood prediction as an image-to-image translation task. By integrating convolutional neural networks (CNNs) with spatial features, such as distance-to-river, this system predicts maximum inundation depths within seconds, generating inundation maps that support real-time risk assessment, traffic management, and emergency response planning. The Feature-Informed Forecast System achieves accuracy comparable to that of hydrodynamic models, thus demonstrating the scalability of data-driven approaches to river flooding in urban areas. Building on this, to capture flood dynamics over time, a Multi-Step Dynamic Forecast System was developed. Employing recursive CNN-based predictions, the Multi-Step Dynamic Forecast System produces temporally resolved inundation maps for up to 24 hours ahead. Compared to physically based models, it maintains high accuracy and reliability, thereby bridging the gap between static maximum flood maps and real-time dynamic monitoring.
Since these and, in general, conventional data-driven models are often case-site specific and therefore domain-dependent, their applicability in different sites is limited. To overcome this constraint, this dissertation proposes a physically informed spatial-temporal forecast framework, in the context of pluvial floods. This approach integrates CNN-based image-to-image translation with physical constraints by embedding the continuity equation and a kinematic wave approximation into the loss function. This framework improves physical plausibility, enhances predictive accuracy, and enables domain-independent forecasts that generalize to new and unknown areas.

Collectively, the contributions of this dissertation demonstrate how data-driven, especially physically informed, models transform flood forecasting into a real-time, scalable, and transferable application. The developed systems advance the state of the art by combining computational efficiency, predictive accuracy, and physical consistency, laying a new foundation for further research in the field of flood forecasting and early warning systems in urban areas.