Estimation of peak outflow in dam failure using neural network approach under uncertainty analysis

This paper presents two Artificial Neural Network (ANN) based models for the prediction of peak outflow from breached embankment dams using two effective parameters including height and volume of water behind the dam at the time of failure. Estimation of optimal weights and biases in the training phase of the ANN is analysed by two different algorithms including Levenberg—Marquardt (LM) as a standard technique used to solve nonlinear least squares problems and Imperialist Competitive Algorithm (ICA) as a new evolutionary algorithm in the evolutionary computation field. Comparison of the obtained results with those of the conventional approach based on regression analysis shows a better performance of the ANN model trained with ICA. Investigation on the uncertainty band of the models indicated that LM predictions have the least uncertainty band whilst ICA’s have the lowest mean prediction error. More analysis on the models’ uncertainty is conducted by a Monte Carlo simulation in which 1000 randomly generated sets of input data are sampled from the database of historical dam failures. The result of 1000 ANN models which have been analysed with three statistical measures including p-factor, d-factor, and DDR confirms that LM predictions have more limited uncertainty band.