Designing High-Precision Fast Nonlinear Dam Neuro-Modelers and Comparison with Finite-Element Analysis

• Published in: Journal of engineering mechanics 2013-10, Vol.139 (10), p.1311-1324
• Main Authors: Joghataie, Abdolreza, Dizaji, Mehrdad Shafiei
• Format: Article
• Language: English
• Subjects: Technical Papers
• ISSN: 0733-9399; 1943-7889
• DOI: 10.1061/(ASCE)EM.1943-7889.0000572

AbstractIn this paper, a method has been proposed to use artificial neural networks for the modeling of concrete gravity dams with nonlinear hysteretic response under earthquake loading. The main advantage of this method is that it makes it possible to design an analysis tool for a specific dam based on the data obtained from monitoring said dam; hence, it is expected that the analysis tool could provide more precise results than analysis software presently available. This advantage is especially pronounced when the dam is to be analyzed under strong earthquakes where its response is nonlinear and hysteretic. The modeler neural network, referred to here as the neuro-modeler, offers considerable possibilities. For example, the nonlinear analysis of dams under earthquake loading requires considerable time; one advantage of designing neuro-modelers for dams is that they can give practically precise results about the response in a short time. As the first study on the subject to prepare a controlled precise data for the training and testing of the neuro-modeler so that the precision of the method could be evaluated, an analysis software was used to simulate the experiment. The smeared crack model, which has been one of the models used successfully in the literature to model concrete gravity dams, has been used in this study as well. Multilayer feed-forward neural networks have also been used. The first step in this method is to analyze the dam under study as it is subjected to different earthquake simulations to collect large data about its linear and nonlinear response. The second step is to train a neuro-modeler, based on the collected data, to implicitly learn the nonlinear hysteretic response of the dam being subjected to the training earthquakes. The third step is to test the precision and generalization capabilities of the neuro-modeler where it is used for the analysis of the dam under a number of selected earthquakes of different properties including both near- and far-field excitations. After passing the tests successfully, the neuro-modeler is expected to be able to provide reliable and precise results about the response of the dam under any given earthquake. Three concrete gravity dams with differing characteristics and in different geographical locations including the Koyna, Pine Flat, and Sefid-Rud dams, which have been both numerically and experimentally studied because they had experienced considerable damage during earthquakes, have been utiliz