基于深度学习算法的地震动重要持时预测模型

Models for predicting ground motion significant duration based on deep learning algorithm

地震动持时对工程结构地震反应有重要影响,发展可靠的地震动持时预测模型是地震工程一个重要课题。基于日本K-NET和KiK-net在1997—2021年间获得的67813条地震动记录,将地震动记录按照震源类型分为浅地壳地震、俯冲带板间地震、俯冲带板内地震和上地幔地震四种类型,采用深度学习算法,建立了四种震源类型地震的地震动重要持时预测模型,并与传统预测方程进行了对比。结果表明,采用深度学习算法建立的地震动重要持时预测模型具有合理性和可靠性,能够取得良好的预测效果。四种类型地震的地震动重要持时的预测结果差异明显,尤其在震级较大的情况下,所以预测不同类型地震的地震动持时不应采用同一种预测模型。研究结果和结论可供地震动参数预测、地震区划、结构抗震设计和地震危险性分析等工作参考。

Ground motion duration has an important impact on seismic response of engineering structures,the development of reliable ground motion duration prediction models is an important topic in earthquake engineering.Here,based on 67813 ground motion records obtained by K-NET and KiK-net in Japan from 1997 to 2021,these ground motion records were divided into 4 types according to source types of shallow crustal earthquake,subduction interface earthquake,subduction slab earthquake and upper mantle earthquake.Using the deep learning algorithm,4 ground motion significant duration prediction models for the 4 source types of earthquakes were established and compared with the traditional prediction equation.The results showed that ground motion significant duration prediction models established using deep learning algorithm are reasonable and reliable,they can obtain good prediction results;ground motion significant duration prediction results for the 4 types of earthquakes have obvious differences,especially,in cases of larger earthquake magnitudes,so the same prediction model should not be used to predict ground motion significant durations for different types of earthquakes;the study results and conclusions can provide a reference for ground motion parametric prediction,seismic zoning,structural aseismic design and seismic hazard analysis.