PhD Oral Exam - Mohamed Ahmed, Civil Engineering

When studying for a doctoral degree (PhD), candidates submit a thesis that provides a critical review of the current state of knowledge of the thesis subject as well as the student’s own contributions to the subject. The distinguishing criterion of doctoral graduate research is a significant and original contribution to knowledge.

Once accepted, the candidate presents the thesis orally. This oral exam is open to the public.

Abstract

Reliable predictions of river flow are of fundamental importance to the sound management of river floods, safe and cost-effective design of infrastructures, and efficient operations of water systems. Despite significant advances in river flow modeling, challenges remain due to uncertainties in model parameters, limited availability of observations, and the inherent complexity of river hydraulics under varying flow conditions. This thesis introduces a novel suite of data-driven and hybrid data assimilation (DA) methods that synergistically integrate hydrodynamic principles with observational data to improve the accuracy and robustness of the prediction methods.

The first part of this thesis is a contribution to the development of a Modified Group Method of Data Handling (MGMDH), a lightweight machine learning model tailored for predictions of the river discharge at a cross-section with limited or missing measurements. MGMDH innovatively incorporates discharge observations from a neighboring hydrometric station, along with meteorological input such as rainfall and temperature, enabling accurate predictions even when direct observations are unavailable. The automatic equation selection mechanism ensures high computational efficiency while maintaining physical interpretability and predictive strength (coefficient of correlation R² > 0.977), outperforming traditional machine learning models.

The second part of the thesis proposes an adaptive Proportional-Integral-Derivative (PID)-based DA framework that dynamically integrates real-time observations of water level and flow velocity into a hydrodynamic model without requiring manual calibrations. This approach addresses a critical limitation in conventional hydrodynamic models—manual tuning of uncertain parameters such as friction coefficient—by using a control-theory-based feedback mechanism to correct the model state during simulation runtime. The PID-DA framework was tested using laboratory data from flume experiments and field data from the Danube River. The framework achieved superior accuracy and efficiency, with relative errors below 1%, and significantly reduced computational costs compared to model predictive control (MPC) based DA. Furthermore, a sequential observation-site selection algorithm grounded in open-channel flow theories identifies optimal assimilation points, enhancing the framework’s spatial adaptability.

The third part of this thesis extends the PID-DA method into a self-adaptive, zone-based framework that calibrates spatially distributed friction coefficients (ks) across highly variable and nonlinear flow regimes. This synergetic approach eliminates the need for linearity assumptions and covariance matrix computations, which are constraints of Kalman Filter-based methods, and allows for rapid, localized model corrections in data-sparse regions. The method was implemented in a five-zone flume testbed. It demonstrates high fidelity (R² > 0.999, σₙ < 1.4%) across both assimilated and non-assimilated nodes, while the Kalman Filter struggles in turbulent or structurally mismatched regions, showing errors exceeding 50%. Comparisons of the predicted velocity field with observational data highlight the critical role of assimilating multi-dimensional flow data, revealing significant improvements over water-level-only DA. The above-mentioned three parts of this thesis have advanced real-time predictions of river flow by offering a spectrum of adaptive, computationally efficient tools that are physically consistent, robust against uncertainties, and scalable to natural rivers. The methods open new directions for integrating remote sensing, in-situ sensing networks, and control algorithms into operational hydro-system management. The new framework provides strong foundations for future smart water infrastructures. It enables faster, more accurate, and more resilient flood predictions and river regulation systems in a changing climate and expanding data environment.