Exploring the Use of Artificial Neural Networks for Scour Prediction

Abstract

Accurate predictions of equilibrium scour depth are of great importance for the design of bridge pier foundations. In recent years, some attempts have been made to develop artificial neural networks (ANN) to predict scour around bridge piers. These ANNs can vary by type [e.g. feed-forward (FF), general regression (GR), radial bias function (RBF)] and the learning algorithm applied [e.g. back propagation (BP), Levenberg–Marquardt (LM), particle swarm optimization (PSO), Firefly (FA)]. The networks also vary by the type of data used (i.e. field or laboratory conditions). The performances of the ANNs have been evaluated by statistical measures such as the coefficient of determination and mean absolute error. The selection of data sets can influence the performance of an ANN; as the variation in the data increases, the accuracy of the output decreases. Most of the literature reviewed in this paper utilized laboratory data to develop the ANNs; although these were accurate in predicting other laboratory data, the accuracy of the predictions when applied to field conditions is unknown. From the ANNs reviewed, it was found that FF-ANNs trained by the PSO algorithm generally yielded the most accurate results. Comparison of the ANN performances to the performances of accepted empirical methods indicates that ANNs produced more accurate results.