New Proposed Model for Predicting Earthquakes Details Using Bi-LSTM

Abstract

An earthquake can be defined as a shaking event that occurs when the tectonic plates of Earth move. Significant harm, including fatalities, structural destruction, and economic effects, may result from such events. The majority of models have only been able to predict certain regions, in spite of multiple attempts to predict such events. For predicting the occurrence and location of earthquakes, this research presents two new models. Bi-directional Long Short-Term Memory (BiLSTM) networks were found to be very appropriate by reviewing the literature because of their efficient memory retention qualities. The best model has been selected by utilizing Keras tuner, which allowed for the selection of different dense layer combinations as well as BiLSTM configurations. The model of choice makes use of seismic markers from earthquake catalog of Bangladesh in order to predict the probability of earthquakes in the future month. An attention process has been incorporated into BiLSTM framework to improve prediction accuracy in the occurrence prediction model, yielding an accuracy rate of 80.1%. Also, an attention mechanism was not included in the location prediction model since it would not improve the performance of BiLSTM architecture and would just add needless complexity. Instead, a regression model has been created by using BiLSTM and dense layers for estimating earthquake epicenter relative to fixed point. Obtaining a root mean square error (RMSE) of 1.1830 as a result.