Ensemble Deep Learning NARX for Estimating Time Series of Earthquake Occurrence

Predicting earthquake occurrences in time series data remains a challenging task in seismology. NARX (Nonlinear Autoregressive with eXogenous input) neural networks have recently shown promise for accurate predictions. While previous research has demonstrated NARX's effectiveness in estimating earthquake frequencies, the architecture still relies on several pre-determined parameters. These parameters include the number of hidden layers, number of neurons, and delay time. This study aims to investigate the effectiveness of Ensemble Deep Learning NARX neural networks in predicting time series of earthquakes. The proposed ensemble model integrates multiple NARX neural networks, each trained on a distinct subset of earthquake data. This approach aims to enhance prediction accuracy and bolster model robustness. The dataset consists of time series records detailing earthquake occurrence frequencies and magnitudes. The results show performance evaluation metrics in terms of Mean Square Error (MSE) values. For training frequency data, the MSE is 4.10e-26, and for testing, it is 6.05e-22. Regarding training magnitude data, the MSE is 2.86e-21 for training and 3.17e-19 for testing. This study makes a valuable contribution to the advancement of earthquake prediction techniques, underscoring the potential of Ensemble Deep Learning NARX neural networks for precise time series estimation in seismology.