Multi-Model Fusion and Multi-Task Learning for Enhanced Earthquake Sequence Prediction Using Acoustic Signals

Accurately predicting earthquake sequences is crucial for mitigating the impacts of seismic events. This study introduces a multi-model fusion and multi-task learning approach to improve earthquake sequence predictions using acoustic signals. By combining LightGBM, support vector regression (SVR), and neural networks trained with multi-task loss functions, the proposed method effectively captures the complex temporal dependencies in seismic data. The workflow includes advanced data preprocessing, signal processing, and feature engineering, followed by model integration using LSTM. Evaluation results demonstrate that the LSTM+LightGBM combination performs best, showcasing the effectiveness of multi-model fusion in earthquake time series prediction. This integrated approach not only addresses the nonlinear nature of seismic signals but also enhances the robustness and accuracy of the predictions compared to traditional single-model methods. The findings provide valuable insights for developing more reliable earthquake forecasting systems, ultimately contributing to better preparedness and risk mitigation strategies.