Abstract

Reliable river flow predictions are vital for flood management, infrastructure design, and efficient water system operations. Despite advances in modeling, challenges persist due to parameter uncertainties, limited observations, and the complexity of river hydraulics. This thesis introduces a suite of data-driven and hybrid data assimilation (DA) methods that integrate hydrodynamic principles with observations to improve prediction accuracy and robustness. The first contribution is a Modified Group Method of Data Handling (MGMDH), a lightweight machine learning model for discharge prediction at sites with limited or missing data. By incorporating discharge from neighboring stations and meteorological inputs such as rainfall and temperature, MGMDH achieves accurate forecasts even without direct observations. Its automatic equation selection ensures efficiency and interpretability, outperforming conventional ML models (R² > 0.977). The second contribution is an adaptive PID-based DA framework, which integrates real-time water level and velocity observations into hydrodynamic models without manual calibration. Unlike traditional methods requiring parameter tuning, the control-theory feedback mechanism dynamically corrects model states. Tested on flume experiments and Danube River data, the framework achieved relative errors below 1% with reduced computational cost compared to model predictive control. A sequential site-selection algorithm based on open-channel flow theory further enhances adaptability by identifying optimal assimilation points. The third contribution extends this into a zone-based PID-DA framework for calibrating spatially distributed friction coefficients across nonlinear flow regimes. This method avoids linearity assumptions and covariance computations of Kalman Filter techniques, enabling rapid, localized corrections in data-sparse regions. Implemented in a five-zone flume, it achieved very high fidelity (R² > 0.999, σₙ < 1.4%) across assimilated and non-assimilated nodes, while Kalman Filter methods showed errors exceeding 50% in turbulent or mismatched regions. Assimilating multi-dimensional flow data further improved velocity field predictions compared to water-level-only DA. Collectively, these contributions advance real-time river flow prediction by providing adaptive, efficient tools that are physically consistent, robust to uncertainties, and scalable to natural rivers. The frameworks support integration of remote sensing, in-situ networks, and control algorithms, enabling faster, more accurate, and resilient flood prediction and river regulation in a changing climate.