基于LSTM网络的单台仪器地震烈度预测模型

Prediction of instrumental intensity for a single station using a LSTM neural network

烈度是地震预警系统的关键产出.如何实现快速预测目标场址的地震烈度是地震预警方法技术研究中的核心问题.本文提出了一种基于长短时记忆神经网络(Long Short-Term Memory,LSTM)的单台仪器地震烈度的预测模型(LSTM-Ⅰ).该模型以一个台站观测到地震动参数的时间序列特征为输入,实现动态预测该台站可能遭受的最大烈度.选取了日本K-NET台网记录的102次地震的5103条强震加速度记录训练了神经网络,利用89次地震的3781条数据检验了模型的泛化能力.利用准确率、漏报率以及误报率三个评价指标评价了LSTM-Ⅰ模型的性能.结果表明,当采用P波触发后3 s的序列进行预测时,模型出现漏报的概率为46.78%,出现误报的概率为1.25%;当采用P波触发后10 s的序列进行预测时,模型出现漏报的概率大幅降低到17.6%,出现误报的概率降低到1.14%.结果表明LSTM-Ⅰ模型很好把握住了时间序列中蕴含的特征.进一步基于LSTM-Ⅰ模型评估了Ⅵ度下台站所能提供的预警时间.本文模型能够提供的预警时间与P-S波到时差接近,说明LSTM-Ⅰ模型具有较高的时效性.

Seismic intensity is a key output of earthquake early warning(EEW)system.How to quickly predict the seismic intensity at the target site is a core issue of EEW related techniques.This paper proposes a Long Short-Term Memory(LSTM)neural network-based instrumental seismic intensity prediction model(LSTM-Ⅰ).The model takes the sequential features of ground motion parameters observed at a station as model input,and continuously predicts the potential maximum intensity at the station.The model was trained using 5103 strong ground motion acceleration recordings of 102 earthquakes from K-NET in Japan.The generalization of the model was tested with the 3781 recordings of 89 earthquakes.The performance of the LSTM-Ⅰmodel is evaluated by three metrics:precision rate,false alarm rate and missed alarm rate.The results indicate that if the first 3.0 s time series segment after the P wave onsets is employed,the probability of the missed alarm is 46.78%,and the probability of false alarm is 1.25%;While the first 10.0 s time series segment is employed,the probability of missed alarm significantly reduces to 17.6%,and the probability of false alarm reduces to 1.14%.The results verify that LSTM-Ⅰmodel has well captured hidden characteristic of the ground motion sequential time series.Furthermore,the lead-time at station with potentialⅥdegree intensity is evaluated,and the lead-time is close to that of P-S wave arrival time difference,which indicates high timeliness of LSTM-Ⅰmodel.