基于机器学习预测模型的现地警报级别地震预警试验——以2022年9月5日四川泸定6.8级地震为例  

Test of on-site alert-level earthquake early warning based on machine learning prediction models

作　　者：宋晋东[1,2] 朱景宝 李水龙 王士成 韦永祥[3] 李山有[1,2] SONG JinDong;ZHU JingBao;LI ShuiLong;WANG ShiCheng;WEI YongXiang;LI ShanYou(Key Laboratory of Earthquake Engineering and Engineering Vibration,Institute of Engineering Mechanics,China Earthquake Administration,Harbin 150080,China;Key Laboratory of Earthquake Disaster Mitigation,Ministry of Emergency Management,Harbin 150080,China;Fujian Earthquake Agency,Fuzhou 350003,China)

机构地区：[1]中国地震局工程力学研究所,地震工程与工程振动重点实验室,哈尔滨150080 [2]地震灾害防治应急管理部重点实验室,哈尔滨150080 [3]福建省地震局,福州350003

出　　处：《地球物理学报》2024年第8期3004-3016,共13页Chinese Journal of Geophysics

基　　金：中国地震局工程力学研究所基本科研业务费专项(2021B07);国家自然科学基金项目(U2039209、42304074和51408564);国家重点研发计划项目(2018YFC1504003)资助。

摘　　要：2022年9月5日12时52分四川省甘孜州泸定县发生6.8级地震,造成严重的经济损失和人员伤亡.本文利用此次地震中台站记录到的强震动数据,离线模拟基于机器学习预测模型的现地警报级别地震预警方法.该方法首先构建基于支持向量机的震级预测模型与现地地震动速度峰值(peak ground velocity,PGV)预测模型,而后将每个台站的震级和PGV预测值分别与震级阈值5.7和PGV阈值9.12 cm·s^(-1)做比较,进而得到现地警报级别(0,1,2,3),并用于判断台站附近是否发生潜在破坏.其中,警报级别3为预测震级和预测PGV都超过了阈值,表明在该台站附近有潜在地震破坏且震级偏大.此次地震的离线模拟结果表明:使用P波到达后3 s时间窗,基于支持向量机震级预测模型的单台震级估计标准差为0.35、平均绝对误差为0.27;基于支持向量机PGV预测模型的现地PGV预测标准差为0.34、平均绝对误差为0.32;震级估计误差和PGV预测误差主要分布在±2倍标准差范围内.在不考虑数据打包与传输延时的条件下,地震烈度Ⅶ度区域内的触发台站在震后8 s几乎都发布了警报级别3.在此次地震的震后初期,基于机器学习预测模型的现地警报级别地震预警方法可以得到可靠的警报预测结果,并为中国地震预警系统升级提供了潜在参考.At 12:52 on September 5,2022,a M6.8 earthquake occurred in Luding County,GarzêPrefecture,Sichuan Province,causing serious economic losses and casualties.This paper uses the strong-motion data recorded by the station for this earthquake to off-line simulate the on-site alert-level earthquake early warning(EEW)method based on machine learning prediction models.The method first constructs magnitude estimation model and on-site peak ground velocity(PGV)prediction model based on the support vector machine(SVM),and then compares the predicted magnitude and PGV of each station with the magnitude threshold(M=5.7)and the PGV threshold(PGV=9.12cm·s^(-1))respectively,so as to obtain the on-site alert levels(0,1,2,3),which are used to judge whether there is potential damage near the station.The alert level 3 denotes that the predicted magnitude and PGV both exceed the thresholds,indicating that there is earthquake potential damage near the station and the event is high magnitude.The off-line simulation results show that using 3 s P-wave time window,the standard deviation of magnitude prediction error of a single station based on SVM magnitude estimation model is 0.35 and the mean absolute error is 0.27;the standard deviation of PGV prediction error based on SVM PGV prediction model is 0.34,and the mean absolute error is 0.32;the magnitude prediction error and PGV prediction error are mainly distributed within±2 times of the standard deviation.Without considering the data packaging and transmission delay,almost all triggered stations in the seismic intensity VII area issued alert level 3 at 8 s after the earthquake occurs.The on-site alert-level EEW method based on machine learning prediction models can obtain reliable alarms at the initial stage after the earthquake occurs,which provides a potential reference for the upgrading of Chinese EEW system.

关 键 词：现地地震预警 机器学习 震级预测 PGV预测 泸定地震 

分 类 号：P315[天文地球—地震学]