Short-term Earthquake Prediction

Summary

Discover how we have performed short-term earthquake predictions with deep neural networks in New Zealand. Since Earthquakes can have tremendous effects, it can be of incredible value to predict earthquakes as early as possible and with the highest accuracy possible.

This project uses a Long-Short Term Memory Network deep neural network architecture. Considerable amounts of data are required to train such an LSTM properly. This data is retrieved from the FDSN dataset. The FDSN dataset is an open-source dataset consisting of seismic measurements of stations around New Zealand and corresponding information when an earthquake happens.

This project was a Bachelor's Thesis for the Computer Science and Engineering degree at TU Delft from Glenn van den Belt.

‍

What we've done

1. Determine the most suitable deep learning method for short-term earthquake prediction.

2. Obtain data for training and testing purposes of the deep learning model.

3. Preprocessing of the dataset; normalisation, down-sampling and sanitisation have been performed.

4. Training the dataset using DelftBlue supercomputer.

5. Test the predictions on the test dataset. The highest accuracy on binary predictions was 64%.

‍

Reference

The TU Delft has approved this research and adopted it into their repository, which is accessible to anyone.

The full project and research paper can be found on https://repository.tudelft.nl/islandora/object/uuid:902c0256-0e4e-410d-8f22-04bf526484d6?collection=education

‍

‍