Attention-based Residual Long Short-Term Memory for Earthquake Return Period Prediction in the Sulawesi Region
Dublin Core
Title
Attention-based Residual Long Short-Term Memory for Earthquake Return Period Prediction in the Sulawesi Region
Subject
earthquake prediction, LSTM, return period, time series analysis, Sulawesi
Description
Indonesia, particularly the Sulawesi region, experiences significant seismic activity due to its position at the convergence of three major tectonic plates. This study seeks to construct a model for predicting earthquake return periods in the Sulawesi area by employing the Residual Long Short-Term Memory (Residual LSTM) architecture integrated with an attention mechanism. The dataset utilized originates from the United States Geological Survey (USGS), focusing on the Sulawesi Island region within the coordinates of latitude -6.184° to 2.021° and longitude 118.433° to 125.552°, spanning the years 1975 to 2024. The research methodology is structured into three primary phases: (1) data collection and
preprocessing, including data cleaning, missing value handling, and normalization, (2) exploratory data analysis to understand seismic data characteristics, and (3) development of the Residual LSTM model with an attention mechanism. The evaluation results show excellent model performance with Train Loss 0.0090, Test Loss 0.0091, Training MAE 0.0698, Testing MAE 0.0717, Training RMSE 0.0947, Testing RMSE 0.0951, and stable Huber Loss of 0.0045 for both training and testing data. The implementation of residual connections successfully addressed the vanishing gradient problem, while the attention
mechanism enhanced prediction interpretability. The small discrepancy between the training and testing metrics confirms the model's robust generalization ability, indicating its strong potential for applications in predicting earthquake return periods.
preprocessing, including data cleaning, missing value handling, and normalization, (2) exploratory data analysis to understand seismic data characteristics, and (3) development of the Residual LSTM model with an attention mechanism. The evaluation results show excellent model performance with Train Loss 0.0090, Test Loss 0.0091, Training MAE 0.0698, Testing MAE 0.0717, Training RMSE 0.0947, Testing RMSE 0.0951, and stable Huber Loss of 0.0045 for both training and testing data. The implementation of residual connections successfully addressed the vanishing gradient problem, while the attention
mechanism enhanced prediction interpretability. The small discrepancy between the training and testing metrics confirms the model's robust generalization ability, indicating its strong potential for applications in predicting earthquake return periods.
Creator
Muhdad Alfaris Bachmid, Daniel Febrian Sengkey, Fabian Johanes Manoppo
Source
DOI: http://dx.doi.org/10.21609/jiki.v18i2.1506
Publisher
Faculty of Computer Science UI
Date
2025-06-26
Contributor
Sri Wahyuni
Rights
ISSN : 2502-9274
Format
PDF
Language
English
Type
Text
Files
Collection
Citation
Muhdad Alfaris Bachmid, Daniel Febrian Sengkey, Fabian Johanes Manoppo, “Attention-based Residual Long Short-Term Memory for Earthquake Return Period Prediction in the Sulawesi Region,” Repository Horizon University Indonesia, accessed January 11, 2026, https://repository.horizon.ac.id/items/show/9878.