Mulitvariate earthquake forecasting

• Main Author: Vaidyanathan Abhishek
• Other Authors: Jagath C Rajapakse
• Format: Final Year Project
• Language: English
• Published: Nanyang Technological University 2023
• Subjects: Engineering::Computer science and engineering
• Online Access: https://hdl.handle.net/10356/165934
• Institution: Nanyang Technological University

Advancements in forecasting methods have led to the development of better learning algorithms. These algorithms vary from using the dependencies across time as well as incorporating exogenous features in forecasting the target variable. Even with these improvements, earthquake prediction has been a challenging task due to the volume and complexity of data. State of the art models in the field of time series forecasting, include SCINet and Informer models. These models are specifically trained to capture complex dependencies across different aspects in time series data. In addition, feature selection methods using neural network based architectures have also improved the forecasting performance. In this paper, we propose a novel architecture using the train mechanisms of SCINet and Informer models, for earthquake prediction. These forecasting models are used as downstream models coupled with the use of network pruning architectures for feature selection. Experiments showed that the train mechanisms of the two models focussed on specific aspects of the time series, generating useful feature representation of the input. The architecture was further tested by comparing against the use of other traditional feature selection techniques under different train settings. Additional experiments analyzed the forecasting performance using a combination of the above models and results showed significant improvement in predictions. Notably, analysis showed that in the scenario of earthquake detection model losses cannot be used as a singular metric to decide on the best model. Certain experiments show significant improvement in forecasting results even with a slight dip in losses. Improvement in computational resources can be used to test out deeper and more complex architecture along with better representation of such complex data.