Earthquake magnitude prediction with long-short term memory and adaptive neuro fuzzy inference system

There may be losses caused by an earthquake, which is one of the natural disasters that occurs suddenly and has a high destructive activity. In order to prevent these losses and take precautions, prediction of earthquake has been a subject of study. Predicting when, where and how many magnitude an earthquake will occur is the study of earthquake prediction. Two models are proposed in this thesis. These are the long-short-term memory networks (LSTM), which gives successful results in prediction problems thanks to the memory mechanism it contains, and the adaptive neuro-fuzzy inference system (ANFIS), which gives successful results in prediction problems with its rule-based structure. The earthquake data of the last 20 years were collected by filtering. Data were processed by feature extraction methods, which is called seismicity indicators in the literature. The calculations were made according to the threshold magnitude selected as 4.0, 4.5 and 5.0, and 3 new datasets were obtained. To be used in the models, the data normalization process was applied so that the values are in the range of 0-1. A shuffling method was applied to the dataset to ensure randomness. The dataset is divided into two, 80% for model training and 20% for model testing. In order to evaluate the performance of the models, the k-fold cross validation technique was also applied. In both models, 6 scenarios were created to use these datasets. Principal component analysis (PCA) was applied to datasets to reduce data dimensionality and the number of features was reduced from 7 to 4. In this way, 3 new datasets were derived. These datasets were also processed in the same way and 6 new scenarios were created. In these scenarios, RMSE and MSE criteria were calculated to evaluate the error rates of the models, and the R2 score was calculated to verify their performance. The LSTM model showed the best performance with an R2 score of 0.99565 in the scenario where the threshold size of the last 20 years data set was selected as 4.0 and the PCA applied data was used. This model will help predict earthquake magnitude by producing consistent results in prediction of earthquake magnitude.