Mohammad Vaghefi

This work experimentally investigated flow patterns along a 180 degree sharp bend with and without a T-shaped spur dike. Acoustic Doppler Velocimetry (ADV) was used to measure the velocity components in x, y and z directions. The main purpose of this study is to employ artificial neural network methods (ANNs) to establish nonlinear models mapping the velocity components into geometric features of the channel (angle of horizon, distance from outside of the bend, and distance from the bed). Discussed ANNs methods are Feed-forward Neural Network (FFNN), Cascade Feed-Forward Neural Network (CFFNN) and Extreme Learning Machines (ELM). Outliers (abnormalities or anomalies) may be produced by different causes in the measured velocities. Before modeling by ANNs to obtain more accurate models, the clustering method using self-organizing maps (SOM) was used to discover and remove velocity data points which deviate from a normal behavior. Performance of the ANN models was evaluated using correlation coefficient (R) and root mean squared error (RMSE). In order to compare the performance of ANNs in velocity modeling at different locations of the bend, seven different data groups were considered using different combinations of samples by random sampling. Comparison of the results of ANN models with the experimental data indicates that in most cases ANNs provide reasonable results and may be employed successfully in estimating velocity in sharp bends with and without the presence of a spur dike. Also, it may generally be concluded that the ANN models without spur dikes are more accurate than those with spur dikes.