A location-dependent earthquake prediction using recurrent neural network algorithms

Abstract

In this paper, we propose a location-dependent earthquake prediction based on recurrent neural network algorithms. The location-dependent prediction consists of clustering the seismic dataset based on its geographical parameters (longitude and latitude) using the K-Means algorithm. In addition, we divide each cluster into two specific subsets. The first set consists of events with an average magnitude between 2 and 5, and the second set contains seismic events with magnitudes >5. Clustering allows our models to focus on each region independently. Associated with splitting methods, it helps them to learn the specific trends of each cluster with accurate performance. Moreover, in our work, the large earthquakes with few events are trained independently and remain unaffected by the other dominant events. We implement our models using the powerful recurrent neural network algorithms: Long Short Memory (LSTM), Gated Recurrent Network (GRU), and their hybrid model (LSTM-GRU). They are tested using data related to three different regions: Morocco, Japan, and Turkey. We evaluate the performance according to three metrics: mean absolute error, mean squared error, and root-mean-square error. Compared to other works in the literature, our models generally perform well, especially in predicting large earthquakes.