ESTIMATION OF PGA USING ARTIFICIAL NEURAL NETWORK

Abstract

The ground motion record at a particular site during a particular earthquake is a result of a complex— nonlinear combination of many factors, There has been a lot of study about deriving a suitable relation to predict strong ground motion more reliably. But little agreement has been reached in this issue. This is more because the relations not only depends upon data selection, characterization of source, path or site or the regression technique employed but also on the purpose for which the equation is intended to be used. So the process of determining the regression relation much depends upon the appropriate judgment of the scholar. In this study, an attempt has been made to develop ANN models to predict peak ground acceleration with the help of Kyoshin Net (K-NET) database from Japan. The data consists of 84,456 time histories with magnitude ranging from 5 to 8 in MJMA scale. For investigating the problem, total eight datasets were prepared based on maximum value of peak ground acceleration, format of input and number of inputs. The basic inputs were magnitude (M), hypocentral distance (d, average shear wave velocity (Vs), average primary wave velocity (VP). average Standard Penetration Test (SPT) blow count (N) and average density (p) along with observed value of PGA. The models were trained for 80% of the data while 10% data was used for cross validation and the rest 10% for testing the model. The adequacy of the models was evaluated on the basis of absolute mean error, percentage error and correlation coefficient. It was observed that, depending upon the method of processing the data, ANN model could predict values with less than 3% error for as high as 48% cases. From the results, it can be affirmed that ANN is an efficient tool for predicting peak ground acceleration.