The Application of Artificial-Neural-Networks for an Efficient Concept for Sediment Management in Rivers

Reservoirs are needed for many important tasks like agricultural irrigation, drinking water supply, flood protection, and energy production. For each of the listed demands, sustainable management of the reservoir’s dam to ensure the availability of the designed storage volume of water is of great importance. However, sedimentation of river reservoirs is a crucial issue worldwide, a major problem being the lowering of its storage capacity. In addition, a reduced retention volume due to sedimentation might also increase the danger potential of floods. Thus, it is essential to develop tools for an efficient concept for sediment management.

Numerical models are widely accepted for designing new reservoirs and dams or to develop management strategies at existing dams to counteract sedimentation. Such computational models can accurately represent reality and give reliable predictions of future developments. However, the more accurate and more complex (e.g. 2D or 3D) models require huge computational efforts and long calculation times. Since morphological developments have a long response time of around 10-20 years, to successfully perform simulations a 2D or 3D model requires high-performance-computing (HPC) capabilities. This resource is not commonly available. Furthermore, the achievement of an optimal design or the evaluation of different management strategies sometimes requires multiple simulations with different scenarios. Hence, simplified 1D models were still applied in this field.

To overcome this issue, we present in this work the application of artificial-neural-network (ANN) in the field of river management. We demonstrate the suitability of this approach exemplarily at a real-world site, the Saalach River in Germany, one of the tributaries of the Danube River Basin. For this river, we developed an ANN architecture, which is able to accurately reproduce the results obtained from the 2D hydro-morphological model system TELEMAC-SISYPHE. The trained ANN is then applied to predict future riverbed level changes only based on the available and measurable inputs. This information is very important for hydropower plant owners and authorities.