Abstract：This paper proposes to use GPS time series predicted by deep learning to study earthquake precursors, taking the earthquake of MS6.4 in Menyuan, on January 21, 2016 as an example. To obtain a high-precision GPS time series prediction model, the LSTM neural network is trained with GPS time series of historical non-seismic time series of Menyuan station(QHME), Minle station(GSML) and Gulang station(GSGL) near the epicenter, and then the GPS time series of non-seismic time series and the period of time before the earthquake in the region are predicted retroactively. Through comparative analysis of the predicted time series and the real time series, we find that most indexes of the similarity between the two time series before the earthquake are lower than those of the time series without the earthquake, which indicates that the predicted time series before the earthquake is obviously different from the real time series. Meanwhile, considering the trend anomaly of the time series before the earthquake, the abnormal time period is considered to have occurred. The three stations have multiple abnormal dates in E,N and U directions, and different stations have the same abnormal date. The discovery of abnormal periods and abnormal dates indicates that earthquake precursors have been explored.
CHEN Shanpeng,YIN Ling,LIANG Shiming et al. Application of Deep Learning to Predict GPS Time Series in Exploring Precursors of Menyuan MS6.4 Earthquake[J]. jgg, 2020, 40(12): 1248-1253.
