Application of Neural Network for Prediction of Earthen Dam Peak Breach Outflow, and Breach Time



Numerous models have been developed in the past decades to explain the complicated earth dams' berach phenomena. These have included physical as well as mathematical and computer models. Among the more widely used dam breach computer models over decades is the BREACH model. It is based upon erosion and soil mechanics equations, hydraulic and sediment transport laws. The difficulty in gathering data motivates one to use other powerful methods. In this study a new method has been developed for prediction of peak breach outflow and breach time through Artificial Neural Networks (ANNs). Toward this end, synthetic breach parameters of about 115 dams were developed by BREACH model, and then employed to train and test the neural networks. The performance of the network model is investigated through a change of input parameters. A most efficient and global model for assessing a dam breach potential is presented. Later, the most significant input parameters affecting dam breach are investigated. Best results were found with back propagation neural network using multiple hidden layers. The most compatible structure for breach outflow prediction possesses the correlation coefficients of 0.992 and 0.909 for training and testing, respectively. As for breach time, a structure was obtained with the correlation coefficients of 0.993 and 0.884 for training and testing, respectively. A forecast study was performed for the case of Mollasadra Dam. Comparisons between the artificial neural network results and dam BREACH model were made, the results indicating that neural networks are appropriate for predicting dam breach parameters.