Development of recurrent neural network model for regional earthquake estimation

In this thesis, it is aimed to implement an application for use of Recurrent Neural Networks (RNN) in earthquake prediction. For this purpose, the most common version of RNN, Long-Short Term Memory (LSTM), has been provided for regional earthquake prediction. The implementation of a deep version of these networks on earthquake data, taking into account the success of the LSTM networks on sequential data sets. For this purpose, a five-layer deep UKVH network model was designed. For experimental studies in Turkey on Bingöl center radius set around that center circular areas were formed in earthquake prediction. Earthquakes that occur in a sequential manner in this circular area depending on time have been used in the training of the LSTM network. It is ensured that the trained network can estimate future earthquakes on the same area. Earthquake data were obtained from B.U. Kandilli Observatory and Earthquake Research Institute. A database was created by querying the earthquakes between the radius and year intervals determined around the center point. Circular 50km, 100km and 200km radius are used in the estimation regions. The activities between the magnitude of 2 and 9, which took place between these radius, was used. When the obtained estimation results are evaluated, it is observed that the application proposed in the thesis predicts earthquake estimations with a reasonable accuracy.