EARTHQUAKE STRENGTH PREDICTION USING GENERALIZED SPACE TIME AUTOREGRESSIVE (GSTAR) MODEL AND ARTIFICIAL NEURAL NETWORK (ANN) WITH THREE DIMENSIONAL WEIGHT MATRIX

• Main Author: Ananda, Vira
• Format: Theses
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/81663
• Institution: Institut Teknologi Bandung

Sumatra Island has active faults, volcanic belts, and subduction zones of the Eurasian and Indo-Australian plates in the eastern Indian Ocean. This makes 5 out of 25 earthquake-prone areas in Indonesia located on Sumatra. Earthquakes cannot be prevented, but their impact can be reduced. One way to reduce earthquake damage is by predicting earthquake strength using a spatiotemporal model like Generalized Space Time Autoregressive (GSTAR). The GSTAR model detects linear patterns in the data, while an Artificial Neural Network (ANN) captures nonlinear patterns in the residuals from the GSTAR model. Additionally, because earthquake epicenters have different depths, a three-dimensional location is used to build weight matrices through Hadamard matrix multiplication. The weight matrices used are uniform, inverse distance, and modified matrices, which show the relationships between different locations. This study focuses on predicting earthquake strength from August 2013 to July 2023 in the Nias-Pagai subduction zone and the North-West Sumatra fault zone with Latitude (-4.544?S – 2.175?N) and Longitude (95.545? - 102.656?E). The study results show that the GSTAR(1;1)-ANN model with modified weight matrices provides the best prediction performance, with an average MAPE of 8.27%. The MAPE calculation shows that using ANN improves the model's performance by 88.89%. Finally, contour maps are used to visualize the original data and the predicted earthquake strength for the next five months.