A Novel Embedded Convolution-NARX Approach with an Ensemble Framework Using Multi-Parameter Seismic Indicators for Earthquake Occurrence Prediction

The accurate prediction of earthquake frequency is currently a critical challenge. Several deep learning models have been developed but the sequential processing often limits the ability to capture complex spatiotemporal dynamics. Therefore, this study proposes a novel Ensemble Convolutional NARX (Ens-ConvNARX) framework where the convolutional feature extraction is embedded directly within the NARX feedback loop to enable a more integrated learning approach. The performance of the framework was validated across two tectonically distinct regions, including West Coast Sumatra, Indonesia, and Southern California, USA. It was also benchmarked against five baseline models which were Convolutional Neural Network (CNN), Long Short-Term Memory (LSTM), CNN-LSTM, Random Forest (RF), and XGBoost. The results showed the superiority of the Ens-ConvNARX model which achieved multi-step predictive accuracies up to 88.5% in West Coast Sumatra and 90.2% in Southern California. An ablation study also confirmed the critical role of the Magnitude of Completeness (Mc) filter, as its removal reduced the one-month-ahead prediction accuracy by 50 percentage points. This study presents a robust and generalizable framework for seismic prediction and shows the important synergy between advanced model architecture and rigorous data preprocessing for improving disaster preparedness and response in seismically active regions.