Predicting the principal strong ground motion parameters: A deep learning approach

Abstract

In this research, new models are developed to estimate the three principal time-domain parameters of seismic ground motion. A novel deep learning (DL) approach coupled with artificial neural network (ANN), namely deep neural network (DNN) is employed for predicting the strong ground motion parameters such as peak ground acceleration (PGA), peak ground velocity (PGV) and peak ground displacement (PGD). This robust technique that has extended the applications of conventional neural networks improves learning of complicated and nonlinear features via increasing the number of layers as well as the neurons in each layer. The proposed models are constructed upon the NGA-West2 database provided by PEER (Pacific Earthquake Engineering Research Center). This database is more comprehensive than NGA-West1 which was mainly considered for developing previous artificial intelligence-based prediction models. Therefore, the new models are more reliable and can be used for wider ranges of predictors. The DNN attenuation models yield accurate estimates of the site PGA, PGV and PGD based on earthquake magnitude, rake angle, source to site distance and soil shear wave velocity. In addition, it is shown that the developed models, with correlation coefficients of 0.902, 0.899 and 0.911 (for PGA, PGV and PGD respectively), outperform the existing soft computing models. Furthermore, the average values of error measures such as MAE, MAPE and RMSE for PGA, PGV and PGD equal to 0.456, 0.758 and 0.581 compare favorably with those of previous models.