In the past 50 years many equations have been developed that predict the value of the peak ground acceleration in case of an earthquake. Authors have developed these equations with different purposes (for universal use or specific area, specified magnitude interval etc.) and by using different databases. The current database PF-L [3] contains 3550 records of strong earthquakes from Europe and America and is more complete in comparison with the past databases. The goal of the thesis is to refit constants of 45 selected equations that predict the peak ground acceleration by using the newer PF-L database and compare their predicted values. For fitting the Levenberg–Marquardt algorithm in the Matlab program was used. Initial constant values are chosen from four increasing intervals [-1,1], [-10,10], [-100,100] and [-1000,1000]. Predicted values were compared by using root mean square error. Further to this, the database is split ten times in learning (90%) and testing (10%) sets. With cross validation the new prediction values were checked. Five best equations are drawn at 3 different magnitudes. The conclusion is that the prediction accuracy of most equations is very similar and that functions fitted with cross validation are almost the same as at fitting the whole PF-L database. A the same time we found lower bound of prediction error. In our case it is situated at rooted mean square error value of 0,08.
|
