Evaluating the Performance of Long Short-Term Memory Neural Network in Predicting Peak Ground Acceleration of Earthquakes Using Shrinking P-Wave Data
Description:
Institutional use cases of Earthquake Early Warning Systems (EEWs) estimates of Peak Ground Acceleration (PGA) using P-wave data include automated high-speed transportation systems that can stop, abort airplane landings, prevent additional cars from riding on bridges and tunnels, initiate shutdown of sensitive industrial infrastructure like gas pipelines before peak ground motion arrives, preventing cascading failures. Traditional EEWS systems rely on the evaluation of 3 seconds(s) of P-wave data to make predictions about PGA. However, this paper evaluates the performance of varying durations of P-wave data on predicting PGA using a Long Short-Term Memory (LSTM) Recurrent Neural Network.
The LSTM model was trained, tuned, and evaluated on nineteen (19) moderate to large earthquake events collected in Japan by the National Research Institute for Earth Science and Disaster Prevention (NIED). The methodology involves three independent experiments using 4 seconds, 3 seconds, and 2 seconds of seismic data recorded after the arrival of P-waves to predict PGA. After evaluating the model's performance on unseen accelerograms, the experiments using 4 seconds, 3 seconds, and 2 seconds of P-wave window data achieved average Test Root Mean Square Error (RMSE) of 0.559, 0.500, and 0.497, respectively. Surprisingly, the experiment using 2-second P-wave data achieves average test RMSE, which is 0.6% less than that of the 3-second P-wave experiment. We conclude that LSTM models are as effective in using accelerograms of the first 2 seconds after the arrival of P-wave to predict PGA as 3 seconds. The result of this study has positive implications for the timely prediction of PGA in existing EEWs and the improvement of existing mathematical models that predict PGA.
Session: Earthquake Early Warning Optimization and Efficacy [Poster]
Type: Poster
Date: 4/20/2023
Presentation Time: 08:00 AM (local time)
Presenting Author: John Owusu Duah
Student Presenter: Yes
Invited Presentation:
Authors
John Owusu Duah Presenting Author Corresponding Author jo156@duke.edu Duke University |
|
|
|
|
|
|
|
|
Evaluating the Performance of Long Short-Term Memory Neural Network in Predicting Peak Ground Acceleration of Earthquakes Using Shrinking P-Wave Data
Category
Earthquake Early Warning Optimization and Efficacy