基于LSTM神经网络的现地烈度实时估算模型——以JMA烈度为例  

LSTM neural network-based onsite seismic intensity real-time prediction model:taking the JMA intensity as an example

作　　者：李山有[1,2] 肖莹 卢建旗 谢志南[1,2] 马强 陶冬旺[1,2] LI Shanyou;XIAO Ying;LU Jianqi;XIE Zhinan;MA Qiang;TAO Dongwang(Key Laboratory of Earthquak Engineering and Engineering Vibration.Institute of Engineering Mechanics,China Earthquake Administration,Harbin 150080,China;Key Laboratory of Earthquake Disaster,Mitigation of Emergency Management,Harbin 150080,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,黑龙江哈尔滨150080 [2]地震灾害防治应急管理部重点实验室,黑龙江哈尔滨150080

出　　处：《世界地震工程》2024年第3期37-45,共9页World Earthquake Engineering

基　　金：中国地震局工程力学研究所基本科研业务费专项资助项目(2018B02);国家重点研发计划项目(2018YFC1504004)。

摘　　要：如何快速并且准确估计目标场点烈度是地震预警中的关键问题。常用基于衰减关系的场点烈度估计和基于P波信息的现地烈度估计往往存在大震烈度低估的问题。本文提出了一种基于长短时记忆神经网络(logn short-term memery,LSTM)的现地JMA烈度持续估计模型。该模型以现地观测地震动的能量、能量增长率、地震动卓越周期和震源距作为输入,以该点的最大仪器地震烈度为预测目标。选取了日本K-NET台网记录101次地震数据作为训练集,94次地震数据作为测试集,训练了现地烈度估算LSTM神经网络模型。结果表明:在采用3 s时窗长度的序列进行预测时,高估的比例为1.51%,低估的比例为4.00%;并且,随着时窗长度的增加,高估和低估的比例也在不断降低。模型对高烈度(大于等于4.5度)样本的预测时效性随震源距的增加而增加,对大震远场高烈度区域能提供20 s以上的预警时间。The accurate and timely estimation of ground motion at target sites is a long-term and arduous task in earthquake early warning research.Commonly employed methods in earthquake early warning,such as ground motion prediction equation-based methods and on-site methods,suffer from large uncertainties as a result of the limited amount of information used in estimations.Here,we propose a long short-term memory(LSTM)neural network method for the continuous prediction of the intensity at a single station.The proposed model inputs four sequential features(the energy,energy increase rate,predominant period of seismic waves,and hypocentral distance of the station)and outputs the predicted maximum seismic intensity at the station.The proposed model was trained with records from 101 earthquake events and tested with records from 94 earthquake events.The results show that when using a 3-second sequence length for prediction,the overestimation ratio is 1.51%and the underestimation ratio is 4.00%;Moreover,as the sequence length increases,the proportion of overestimation and underestimation also decreases.The timeliness of the proposed model is evaluated by the lead-time that high intensity(intensity great than 4.5)is correctly predicted before the intensity 4.5 is observed.The result show that the lead-time increases with the increase of hypocentral distance.For large earthquakes,the proposed model can provide at least 20 s for end users located in far-field.

关 键 词：地震预警 现地预警 长短时记忆神经网络 实时减灾 烈度估计 

分 类 号：P315.3[天文地球—地震学] P315.7[天文地球—固体地球物理学] P315.92[天文地球—地球物理学]