Minh Duc Bui, Keivan Kaveh, Peter Rutschmann
Monday 29 june 2015
17:48 - 17:51h at Mississippi (level 1)
Themes: (T) Sediment management and morphodynamics, (ST) Sediment transport mechanisms and modelling, Poster pitches
Parallel session: Poster pitch: 3A. Sediment - Erosion
One of the weakest points of hydomorphological models is to use empirical formulae for calculating sediment transport rates, which are of limited generality. In many cases, unreasonable morphological changes are predicted and the results of the different formulae often vary strongly. The reasons are assumed in the complexity of the interaction between flow and sediment transport and in limitations of the nonlinear regression applied in these empirical methods. In recent years, the methods of artificial neural networks (ANN) provided good results in the fields of hydraulic engineering. In contrast to most traditional empirical methods, which need prior knowledge about the nature of the relationships among the data, ANN learns from data examples presented to them in order to capture the subtle functional relationships among the data even if the underlying relationships are unknown or the physical meaning is difficult to explain. Further, this method does not need to be introduced with an explicit form of the regarded task and additionally has proven a high tolerance against data sample errors. These attributes make the utilization of artificial neural networks for sediment transport predictions very promising. Within the frame of an on-going research project at the Institute of Hydraulic and Water Resources Engineering, Technische Universität München, Germany, a new concept for hydromorphological model in alluvial rivers using data driven methods will be tested and developed. In this paper an optimal ANN model is selected, which could adequately predict the morphological changes in straight alluvial channels under steady flow discharges. For this purpose the capability and accuracy of numerous ANN models designed with different structures and trained with different learning rules is analyzed. To evaluate the prediction qualities of the designed networks, a comparative study is carried out for these models by evaluating several statistical parameters that describe the errors associated with the model in terms of statistical measures of goodness-of-fit between the estimated bed changes and analytical solutions.