臺灣位於板塊交界帶，地震發生頻繁，且常有致災型地震造成重大傷亡及經濟損失。倘若能有效地預測地震災害發生的時間、地點以及規模，將有助於減少其帶來的災損。傳統的地震預測方法通常依賴於識別地震發生的前兆或異常現象，然而此類型的方法難以進行普及化應用。近年來，數據驅動的機器學習方法在預測自然災害方面顯示出了更可靠的結果。然而，現有的研究通常需要複雜的資料處理和地震指標的選擇。為了因應這些挑戰，本研究開發了一種直接利用地震目錄進行地震預測的方法。其中透過滑動視窗法將2002年至2022年的原始歷史地震目錄處理為多元時間歷時數據，並藉由三個基於注意力機制的雙向長短型記憶神經網絡(Bidirectional Long Short-Term Memory, Bi-LSTM)模型，由先前連續發生的地震事件來預測即將發生的地震的時間、規模和位置，最後針對預測模型進行了數值實驗來優化超參數。時間預測是透過基於迴歸的學習開發，而規模和位置的預測是透過基於分類的學習來實現。與先前的模型相比，本研究提出之模型的時間預測獲得更佳的R^2，規模和位置預測得到更佳的F_1分數。儘管由於數據不平衡而容易出現過擬合，但結果顯示了使用簡單方法來加強地震預測能力的潛力。這項研究不僅推進了臺灣的地震預測，也為其他具有類似地震複雜性的地區提供了一個可延伸發揮的模型。;Taiwan is situated at a tectonic boundary, making it highly prone to severe earthquakes that often result in significant loss of life and substantial economic damage. If earthquake characteristics such as the time, location, and scale of earthquake events can be effectively predicted, it will help reduce the damage caused by them. Traditional earthquake prediction methods often rely on identifying precursors or anomalies before the events. However, this type of method is difficult to universally apply. In recent years, data-driven machine-learning approaches have shown more reliable results in predicting natural disasters. However, existing studies usually require complex data processing and the selection of seismic indicators. In response to these challenges, this study introduces a novel approach that directly utilizes the earthquake catalog. By developing three attention-based Bi-LSTM models that process raw historical earthquake data from 2002 to 2022 into multivariate time series data via a sliding window technique, this research aims to predict the time, magnitude, and location of upcoming earthquakes based on previously consecutive events. Time prediction was developed through regression-based learning, while the prediction of magnitude and location was implemented through classification-based learning. Numerical experiments were conducted to optimize hyperparameters, resulting in superior R^2 of the time prediction and F_1 scores for magnitude and location over previous models. Despite some susceptibility to overfitting due to the data imbalance, the results highlight the potential of using a straightforward approach to enhance earthquake prediction capabilities. This study not only advances earthquake prediction in Taiwan but also suggests a scalable model for other regions with similar seismic complexities.
