Publikation
From levels to velocities: Benchmarking machine learning approaches for river current prediction
Felix Becker; Christoph Michael Tholen; Frederic Theodor Stahl
In: International Conference on Knowledge-Based and Intelligent Information & Engineering Systems. International Conference on Knowledge-Based and Intelligent Information & Engineering Systems (KES-2026), September 9-11, Dublin, Ireland, Pages 6561-6570, Vol. 292, Elsevier, Amsterdam, 9/2026.
Zusammenfassung
Accurate measurements of water current velocity are essential for a number of applications, including river navigation safety and real-time decision support in tidal environments. However, direct current velocity observations in tidal rivers are typically spatially sparse, costly to deploy, and difficult to acquire continuously in real-time. Hence, this paper presents an approach to estimate the current velocity of tidal rivers based on machine learning methods, to ascertain their capacity to infer current velocity from changes in the water level. The models are trained on water-level deltas and validated against measured current velocities from multiple stations with differing hydrodynamic conditions. Among all of the evaluated methods, the Transformer model was found to achieve the lowest root mean squared error (RMSE) of 5 cm/s across all monitoring stations evaluated. This corresponds to approximately 4% of the full observed velocity range. The model also exhibited the most robust alignment with tidal extrema, outperforming both classical regressors and recurrent neural network variants. The findings indicate that data-driven models possess the capability to
estimate current velocity with a high degree of reliability, thus obviating the necessity for comprehensive hydrodynamic inputs. This finding suggests a promising avenue for the development of real-time river navigation support systems.
