Home // DATA ANALYTICS 2016, The Fifth International Conference on Data Analytics // View article

Leveraging Analytics to Predict Geomagnetic Storms: Impact to Global Telecommunications

Authors:
Taylor Larkin
Denise McManus

Keywords: geomagnetic storms; stacked generalization; regularization; predictive modeling

Abstract:
Coronal mass ejections are colossal bursts of magnetic field and plasma from the Sun. These eruptions can have disastrous effects on Earth's telecommunication systems and power grid infrastructures costing millions of dollars in damages. Hence, it is imperative to construct intelligent predictive processes to determine whether an incoming coronal mass ejection will produce devastating impacts on Earth. One such process, called "stacked generalization," is an ensemble strategy that incorporates the predictions from a diverse set of models (base-learners) by using them as inputs for another model (a meta-learner). The goal of this meta-learner is to deduce information about the biases from the base-learners and improve generalization to make more accurate predictions. In this work, 30 models are chosen from the R package caret to serve as base-learners in order to predict a geomagnetic storm index value associated for 2,811 coronal mass ejection events that occurred between 1996 and 2014. Two meta-learners are explored: 1) standard linear regression 2) non-negative elastic net regression. Results show that for this dataset, stacked generalization with the latter meta-learner produces the lowest error and performs significantly better than any of the base-learners executed individually. Not only does non-negative elastic net regression have predictive advantages, but it provides sparser solutions and more reliable inferences at the meta-level compared to linear regression. This, in turn, encourages the idea of parsimony and consequently, improves the overall generalization behavior of this technique.

Pages: 8 to 13

Copyright: Copyright (c) IARIA, 2016

Publication date: October 9, 2016

Published in: conference

ISSN: 2308-4464

ISBN: 978-1-61208-510-4

Location: Venice, Italy

Dates: from October 9, 2016 to October 13, 2016