NARX-Based 1D Convolutional Neural Networks for Enhanced Earthquake Prediction Accuracy

Earthquake prediction is a challenge at this time. This is because the characteristics of earthquakes are very complex and dynamic. This study aims to create a new method that integrates the Non-linear Autoregressive with Exogenous Inputs (NARX) model and 1-dimensional Convolutional Neural Network (1D-CNN) to improve the accuracy of predicting the number of earthquake events in one month. We design the NARX architecture, based on 1D-CNN, to predict earthquake time series data from three different locations in Indonesia: Sunda Strait, South Java, and Bali. The training and testing process was carried out to predict the number of earthquake events in the coming month. The testing yielded the Mean Squared Error (MSE) metric, which demonstrates the good performance of the proposed model. The MSE values for each region of the Sunda Strait, South Java, and Bali are 2.130e-05, 6.018e-02, and 2.524e-02, respectively. The Mean Arctangent Absolute Percentage Error (MAAPE) metric at the prediction stage shows high accuracy in the first month, where the model is able to predict earthquakes in the short term. This research is expected to be able to answer the challenges of earthquake prediction in the field of seismology. Future developments use other deep learning methods for earthquake prediction.