基于溶解氡数据和长短期记忆网络的地震预报  

Earthquake Prediction Using Dissolved Radon Data in Long and Short Term Memory Network

作　　者：刘海军 单维锋 耿贵珍 LIU Hai-jun;SHAN Wei-feng;GENG Gui-zhen(School of Emergency Management,Institute of Disaster Prevention,Langfang 065201,China;School of Economics and Management,Institute of Disaster Prevention,Langfang 065201,China)

机构地区：[1]防灾科技学院应急管理学院,三河065201 [2]防灾科技学院经济管理学院,三河065201

出　　处：《科学技术与工程》2020年第10期4029-4035,共7页Science Technology and Engineering

基　　金：中国地震局教师科研基金(20150110)。

摘　　要：溶解气氡浓度异常为可靠地震前兆,通过对历史观测数据进行建模,预测溶解气氡未来趋势,是快速检测溶解气氡浓度异常、研究震-氡机制的前提。溶解气氡浓度数据为典型的时间序列数据,传统的时间序列预测技术主要为自回归(AR)方法和自回归滑动平均(ARMA)方法。这些方法以线性方法为主,其拟合精度有限。采用目前最流行的深度学习技术长短期记忆(LSTM)模型对姑咱地震台、西昌地震台和雅安地震台一段时间内连续观测的溶解气氡日观测数据集溶解气氡浓度数据进行建模,采用90%的数据作为训练数据训练LSTM网络,10%的数据作为预测数据,采用均方根误差评价指标来评价模型的效果。在三种数据集上,LSTM的预测误差均方根误差均明显低于AR和ARMA方法。该结果表明,LSTM的预测精度高于传统的AR、ARMA方法。The abnormal concentration of dissolved gas radon is a reliable earthquake precursor.It is the premise of detecting the abnormal concentration of dissolved gas radon and studying the earthquake-radon interrelationship to predict the trend of dissolved gas radon by modeling historical observation data.The radon concentration data of dissolved gas is a typical time series data.The traditional time series prediction techniques are mainly AR(auto-regressive)method and ARMA(auto-regressive moving average)method.These methods are mainly linear,and their fitting accuracy is limited.The most popular depth-learning technology LSTM(long-short-term memory)was used to model the daily observation data set of dissolved gas radon of Guzan,Xichang,and Ya’an seismic stations for a period of time.90%of the data was used as training data to train LSTM network,10%of the data were used as prediction data,and RMSE(root-mean-square-error)evaluation index was used to evaluate the effect of the model.The results show that RMSE was significantly lower than those of AR and ARMA,and the prediction accuracy of LSTM was higher than those of traditional AR and ARMA.

关 键 词：时间序列分析 长短期记忆网络 前兆数据 趋势预测 循环神经网络 

分 类 号：TP29[自动化与计算机技术—检测技术与自动化装置]