Application of Artificial Neural Network Methods to Anatolian Plate Earthquake Magnitude and Location Prediction

Abstract

Same-region earthquakes usually have a pattern that is difficult to identify clearly. Therefore, time series analysis methods have been proposed for earthquake prediction. Our work attempts to predict three earthquake parameters in the Anatolian Peninsula using pure artificial neural network methods. An optimized BP-NN model and optimally hyper-parameterized LSTM Model have been designed to predict earthquake magnitude, latitude, and longitude. The two models are compared with previous works for their prediction performances using four well-accepted metrics: mean squared error, mean absolute error, median absolute error, and standard deviation. The time, depth, sun, and moon distances to Earth were identified as the most contributing factors in earthquake occurrence through analysis by five different feature extraction algorithms. The date harmed the prediction accuracy. The LSTM model outperformed the BP-NN Model in magnitude prediction with 0.062 MSE. Latitude predictions of both methods were satisfactory and close. However, BP-NN had lower error rates in latitude prediction. However, longitude prediction errors were significant in both models. Therefore, our designs did not successfully predict the exact location of the earthquakes. However, multi-variate, stacked LSTM models are promising in predicting Anatolian Peninsula earthquake magnitudes, but future work is necessary for location and timing predictions.