A Deep Learning Framework for Peak Ground Velocity Prediction Using Multi‐Station Velocity Waveforms: The Taiwan Transformer Shaking Alert Model (TT‐SAM)

This study proposes a deep‐learning–based regional earthquake early warning model, the Taiwan Transformer Shaking Alert Model (TT‐SAM). The model adopts peak ground velocity (PGV) as its primary ground shaking prediction unit, aiming to better reflect actual structural damage and thereby enhance the practical utility and accuracy of the warning system. Unlike most deep learning earthquake warning approaches that rely solely on raw waveforms, TT‐SAM integrates velocity waveforms, low‐frequency velocity waveforms, and physically meaningful seismic features to enhance the model's ability to capture strong ground‐motion characteristics. A rolling‐update prediction strategy enables preliminary forecasts within 3 s of P‐wave arrival, with continuous updates as additional waveform data become available. To simulate conditions of incomplete waveform acquisition, a time‐masking mechanism is applied during training to improve generalization. Furthermore, magnitude oversampling and bias‐to‐close‐station strategies are employed to alleviate underprediction arising from scarce high‐intensity samples in the training data. When evaluated on the Taiwan seismic events data set in 2016, TT‐SAM achieved 97.8% of its intensity‐level predictions within ±1 level, demonstrating outstanding predictive capability and stability. By directly predicting PGV and aligning model outputs with official intensity‐based warning criteria, TT‐SAM offers a robust and practical framework for next‐generation EEW systems. The proposed approach demonstrates the value of combining deep learning with physics‐guided feature engineering to enhance both the accuracy and reliability of real‐time seismic warning.