Performance Analysis Of Different Model Architectures Utilized In An Adaptive Neuro Fuzzy Inference System For Contraction Scour Prediction

The processes involved in the local scour due flow contraction are so complex that it is difficult to establish a general empirical analytical model to provide accurate estimation of scour. In this paper, the capacity of an Adaptive-Network-Based Fuzzy Inference System (ANFIS) for predicting equilibrium contraction scour depth in alluvial channels was investigated. The main subject of this work is to design an appropriate neural network architecture for training the ANFIS from a given set of input and output data. The training algorithms used in this work are (1) basic hybrid method, (2) basic backpropagation with gradient descent method, (3) backpropagation with momentum method, and (4) backpropagation with Levenberg-Marquardt method. Applying a self-developed software, the numerical experiments were carried out by combining these training algorithms with different ANFIS structures. Statistical indices of model performance such as mean average error, root mean squared error, and coefficient of correlation were measured for each combination. The results showed that among all given models the zero order Takagi-Sugeno’s model with four bell-shaped membership functions for each input and the Levenberg-Marquardt algorithm for training provided the best performance for estimating of contraction scour depth.