Estimating Values of the Maximum Peak Ground Acceleration of a Strong Motion by Three Models of Artificial Neural Networks

Peak ground acceleration is one of the most important factors that needs to be investigated in order to predict the devastation potential resulting from earthquakes in reconstruction sites. Besides, the maximum level of shaking control is subjected criteria that can be worth considering. In this research, a training algorithm based on gradient descent and Levenberg-Marquart (Train LM) were developed and employed by using strong ground motion records. The Artificial Neural Networks (ANN) algorithm indicated that the fitting between the predicted PGA values by the networks and the observed PGA values were able to yield high correlation coefficients of 0.78 for PGA.
From a deterministic point of view, the determination of the strongest level of shaking that is expected at a site has long been a significant consideration in earthquake engineering. Besides, knowledge of the maximum physically possible ground motions allows a meaningful truncation of the distribution of ground motion residuals, and as a result, leads to falling of the values computed in probabilistic seismic hazard analysis (Strasser and Bommer, 2009).
The peak ground acceleration parameter is often estimated by the attenuation of relationships and by using regression analysis. PGA is one of the most important parameters, often analyzed in studies related to damages caused by earthquakes (Gullo and Ercelebi, 2007). It is mostly estimated by the attenuation of equations and is developed by a regression analysis of powerful motion data.
Kerh and Chaw (2002) used software calculation techniques to remove the lack of certainties in declining relations. They used the mixed gradient training algorithm of Fletcher-Reeves’ back propagation error (Fletcher and Reeves, 1964). They applied three neural network models with different inputs including epicentric distance, focal depth and magnitude of the earthquakes. These records were trained and then the output results were compared with available nonlinear regression analysis.
In this article, to estimate strong ground motion acceleration component in an area, four artificial neural networks with different algorithms were used, including General Regression Neural Network (GRNN), Nonlinear Auto Regression neural network (NARX), Feed-Forward Back-Propagation error (FFBP) and General Feed-Forward Neural Network (GFFNN). Input vectors of neural networks include four parameters, which have key effects in occurrence of an earthquake in an area. The parameters include magnitude of moment, rupture distance of earthquake center, mechanism of faults, and ranking of site. Output vector has only one component: maximum peak ground acceleration for an earthquake in an area is used as a target output.
After different tests, GRNN network has maximum output correlation coefficient (0.87) and General Feed-forward Back-Propagation error neural network (FFBP) has the least (0.41). Besides, GRNN network had the least mean square error (0/014), and Back-Propagation network had 0.125. In this research, GRNN neural network is the best neural network, which can estimate possible peak acceleration more than 1g in an area.
Artificial neural networks are a set of non-linear optimizer methods which do not need certain mathematical models in order to solve problems. In regression analysis, PGA is calculated as a function of earthquake magnitude, distance from the source of the earthquake to the site under study, local condition of the site and other characteristics that are linked to the earthquake source such as slippery length and reverse, normal or wave propagation. In non-linear regression methods, non-linear relations that exist between input and output parameters are expressed as estimations, through statistical calculations within a specified relationship (Douglas, 2003).