A Real-Time Seismic Intensity Prediction Framework Based on Interpretable Ensemble Learning

Abstract Real-time seismic intensity estimation aims to predict the maximum possible damage caused by an earthquake based on primary waves (P wave), so that the earthquake early warning (EEW) targets can take measures to reduce the potential damage according to the predicted seismic intensity. The peak P-wave displacement amplitude (Pd) is often used as an effective characteristic parameter to predict ground-motion peaks; however, it is difficult to accurately predict the complex nonlinearity between P wave and the peak ground motion using a single parameter. To address this problem, we propose a reliable and efficient real-time seismic intensity prediction framework by investigating and comparing the performance of multiple ensemble learning algorithms using the Kyoshin network (K-NET) dataset, with 52,560 sets of three-component records from 2010 to 2018 as training and test sets, and 9166 sets obtained from 2019 to 2021 as a case study. The proposed framework optimizes the ensemble learning models according to the correlation between characteristic parameters to eliminate redundant and irrelevant parameters. An optimal model with 14 characteristic parameters is determined. In addition, we apply interpretable approaches to explain the effects of different parameters on the results in response to the fact that the poor interpretation of machine learning methods leads to low credibility. We verify the efficiency and prove the generalizability of the model using case sets. The results show that the optimized model can predict the maximum intensity with an accuracy rate exceeding 95% within the 1 s time window after the arrival of P wave, and the accuracy stabilizes at more than 97% after 3 s. The framework established in this study can effectively and continuously predict seismic intensity and provide a potential method for EEW.