Continuous prediction of onsite PGV for earthquake early warning based on least squares support vector machine

In order to improve the accuracy and continuity of the on-site instrumental seismic intensity prediction, studying the PGV continuous prediction model for earthquake early warning. Predicting the 0.1 similar to 10 Hz band-pass filtered three-components vector synthetic peak velocity, the Chinese instrument seismic intensity standard, using the Japanese K-net and KiK-net network strong earthquake data in the 1 similar to 10 s time window after P wave arrivals, based on the machine learning method in artificial intelligence, least squares support vector machine, selecting 7 kinds of feature parameters as input to construct the least squares support vector machine PGV prediction model LSSVM-PGV. The results show that the prediction error standard deviation of the LSSVM-PGV model established in this paper on the training data set and the test data set tends to be consistent, LSSVM-PGV model has generalization performance. The predicted PGV and the measured PGV in 3 s after P wave arrivals can meet the 1 :1 relationship as a whole, as the time window increases, the standard deviation of the PGV prediction error decreases significantly, and tends to converge in 6 s after P wave arrivals, this shows that the LSSVM-PGV model has accurate continuous prediction capabilities. Compared with the common P-d-PGV model that is also the 3 s after P wave arrivals, the standard deviation of the PGV prediction error of the LSSVM-PGV model is significantly reduced, "overestimation on small value" and "underestimation on large value" phenomena have been significantly improved, and prediction accuracy has been improved. The analysis of earthquake examples of the Kumamoto earthquake sequence shows that for earthquakes below M(i)6. 5, the LSSVM-PGV model can predict a PGV that conforms to the 1 :1 relationship with the measured PGV overall at most 3 s after P wave arrivals. For the M(j)7. 3 main shock, due to the complexity of its rupture process, the predicted results in 3 s after P wave arrivals are somewhat underestimated, but as the time window grows to 6 s, the predicted PGV and the measured PGV are in a 1 :1 relationship, and the overall trend remains consistent until 10 s. The LSSVM-PGV model constructed in this paper can be used to predict the instrumental seismic intensity of on-site earthquake early warning.