Combining MRA y LSTM for predicting the sismicity in Chile

The present work is carried out considering the high level of complexity related to the prediction of seismic events, since a good prediction of seismicity would improve the ability to make decisions in advance and thus avoid catastrophic effects. This study focuses on the analysis of seismicity in two areas of Chile and proposes a new methodology to predict the seismic intensity function (expected number of seismic events per day) based on multiresolution wavelet analysis and the Long Short Term Memory (LSTM) neural network. First of all, the intensity function is calculated using the ETAS model (Epidemic Time After Shock Sequences), then the Maximal overlap discrete wavelet decomposition (MODWT) is applied to this function and the multiresolution analysis (MRA) is considered for obtaining a time series at each level of resolution. Finally, a LSTM network is applied to each series for predicting the future behavior. The final prediction of the intensity function is then obtained by the sum of the predictions at each resolution level, according to the multiresolution analysis. Although the methodology requires some improvements in the experimentation stage, it is shown that the prediction of the intensity function improve significantly when the LSTM is applied to the decomposed señal instead to the original function. The proposed metodology provided results with a R 2 greater then 0.95 in the testing set for the two studied zone.