Exploring the reliability of different artificial intelligence techniques in predicting earthquake for Malaysia

Abstract

Earthquakes have been universally recognised as seismological disasters that pose a threat to civilization and need to be monitored through prediction models. The development and usage of traditional statistical predicting models, which require the understanding of underlying physical scientific processes in a system and large amounts of data preparation, can be challenging and costly. Artificial intelligence-based models, specifically machine learning models, are able to easily review mass data volumes and identify complex data trends to make predictions, making them beneficial to be utilized as prediction models. Terengganu, located on the east coast of Peninsular Malaysia, has experienced three earthquakes in the last four decades and has the potential to be hit or affected by earthquakes due to its location within the vicinity of the South China Sea where the seismologically active Manila Trench is situated. This makes the development of machine learning models for the prediction of earthquakes in Terengganu important for future disaster analysis and management. Therefore, this study suggests artificial neural network (ANN) models as a tool to predict ground motion parameters, namely earthquake acceleration, depth, and velocity, in Terengganu. However, this study presents the comparison of the results of ANN with the results of Random Forest (RF). The data used to develop the models were collected by six seismological stations for two channels in Terengganu and provided by the Malaysian Meteorological Department. The data was partitioned into six sets for each ground motion parameter in each channel, with each set utilizing data from a different grouping of five stations for training and one station for testing. Earthquake depth was able to be modelled with accuracy univariately, that is using only the respective output parameter, which is earthquake depth, as the input parameter. Earthquake acceleration and velocity could not be modelled with accuracy univariately, and were improved by adding earthquake depth as an input parameter. Based on the analysis and evaluation of the results using four selected performance criteria, the ANN models show good performance in predicting earthquake acceleration, depth, and velocity.