Earthquake Early Warning is a technology that provides information about earthquakes before strong ground motions, such as S-waves arrive, thereby reducing damage and securing response time. In particular, on-site earthquake early warning predicts gro...
Earthquake Early Warning is a technology that provides information about earthquakes before strong ground motions, such as S-waves arrive, thereby reducing damage and securing response time. In particular, on-site earthquake early warning predicts ground motion at a specific location using earthquakes information recorded at a single station. Previous studies mainly used empirical equations based on quantitative parameters extracted from P-waves, such as displacement and period. However, recently, with the advancement of related technologies, studies based on various machine learning models have been actively conducted.
In this study, a model for predicting Peak Ground Acceleration (PGA) based on initial P-wave information is proposed for on-site earthquake early warning. Three-component acceleration records were extracted from the STEAD dataset, and time-domain and frequency-domain information from the 3-second window after P-wave arrival were used to construct the input data. A model based on a Transformer encoder was designed, and input data configurations were compared under the same data and training conditions. In addition, proposed model was compared with a CNN model used as the baseline model.
Model performance was evaluated using static and dynamic evaluation approaches. In the static evaluation, a fixed 3 second input window was used to assess prediction performance based on
evaluation metrics and accuracy for each PGA range. In the dynamic evaluation, the input length was increased at 0.1-second intervals to examine early-stage accuracy, prediction stabilization and early prediction performance. The results showed that the Transformer model using frequency-domain information achieved the best overall PGA prediction performance in both evaluations. These results indicate the applicability of frequency-domain information from initial P-wave signals to on-site earthquake early warning. In addition, this study provides a systematic evaluation framework by applying both static and dynamic evaluation methods.