Contraction scour estimation using data-driven methods

The paper presents an artificial neural network (ANN) and an adaptive neuro-fuzzy inference system (ANFIS), which could accurately estimate the maximum equilibrium depth of the contraction scour. The developed networks have been trained using the data set conducted by different investigators for long contractions under clear-water conditions. The designed multilayer perceptron (MLP) ANN includes one hidden layer and seven nodes within that layer. Its hidden neurons use a hyperbolic tangent sigmoidal transfer function. The ANN model was implemented using the MATLAB software package. The importance of the individual input parameters was tested with a sensitivity analysis. This revealed the contraction ratio to be by far the most sensitive parameter, followed by the effect of armor layer formation for nonuniform sediments. For the designed MLP–ANN network, the training was based on the Levenberg–Marquardt back-propagation algorithm in batch mode. The designed ANFIS was the zero-order Takagi–Sugeno model with four bell-shaped membership functions for each input and applied the Levenberg–Marquardt algorithm for network training. The ANFIS model was implemented using a FORTRAN-based computer code. The calculated results show that the selected networks estimate the equilibrium maximum scour depth under clear-water conditions within the range of the used data set significantly better than other conventional methods.