| description abstract | Located at a tectonic boundary, Taiwan is highly prone to severe earthquakes that often cause significant loss of life and economic damage, underscoring the need for improved prediction methods. While traditional approaches focus on detecting precursors, data-driven machine learning and deep learning have demonstrated greater reliability. However, existing models frequently require complex data processing and indicator selection. This study introduces a novel approach that directly utilizes earthquake catalogs, and investigates its adaptability in Taiwan. By developing three attention-based bidirectional long short-term memory (Bi-LSTM) models, historical earthquake data from 2002 to 2022 was transformed into a multivariate time series using a sliding window technique. These models predict the time, magnitude, and location of future earthquakes based on consecutive prior events, with time prediction achieved through regression and magnitude and location through classification. Numerical experiments optimized the models’ architecture and hyperparameters, resulting in superior R2 and F1 scores compared to existing studies. Despite challenges such as data imbalance, which affected the accuracy of magnitude and location predictions during testing, this study makes significant contributions by demonstrating how data processing optimizations can effectively enable Bi-LSTM models to be used across regions with varying seismic characteristics. The streamlined approach not only enhances earthquake prediction in Taiwan but also offers a scalable solution for other regions with complex seismic environments. | |
