Real-time prediction of peak ground acceleration for on-site earthquake early warning in China via Interpretable deep learning

Abstract

A Transformer-PGA network was developed for real-time peak ground acceleration (PGA) prediction of on-site earthquake early warning in China. It was designed to reduce magnitude saturation and to correct the underestimation of strong ground motion in current on-site earthquake early warning methods. A dataset of 6181 three-component ground motion records in China was used. Twenty-eight ground motion parameters were extracted from 0.5 to 50 s after P-wave arrival to train the model. Arias intensity, the average of the top ten peak velocities, and P-wave acceleration were identified as the dominant parameters. Arias’ intensity had an average sensitivity index of about 0.4. Three sub-models were derived with 12, 6, and 3 optimized input parameters. The Transformer-PGA model had a STD of 0.569 at 3 s, outperforming the P_d model (0.743) in reducing PGA underestimation. The sub-models retained over 97 % of the full model’s explained variance as measured by R². Compared with the LSTM model, the Transformer-PGA model showed higher prediction accuracy and produced earlier warnings. Typical warning times were 3 s for Transformer-PGA and 5 s for the LSTM model. For operational use, a dual threshold warning scheme is proposed. An initial alert is issued at 3 s to balance timeliness and accuracy. A refined report is issued at 10 s. In extreme cases, 1 s alerts are provided for emergency response. By combining a Transformer architecture with sensitivity-driven parameter optimization, improvements were made in accuracy, timeliness, and interpretability of real-time PGA prediction for earthquake early warning in China.