A comparison of machine-learning techniques for the prediction of the auroral electrojet index

Abstract

The modern ‘machine-learning models’ are a section of artificially intelligent machines used to implement complex models, which can learn and improve from experience with respect to certain class of jobs, without being specifically programmed. In the present analysis, a comparative study is made of the popular machine-learning techniques regarding the prediction of auroral activity as reflected by the auroral electrojet index (AE index) during geomagnetically disturbed periods. The study also explores the suitability of the online sequential version of the best machine-learning algorithm, which has the potential for real-time forecast of the AE index from short-time input datasets with extremely fast convergence than batch-training methods. The study discusses the need for the correct choice of the input dataset that can be used for predicting the AE index from several combinations of input datasets which include coupling functions, geomagnetic indices and solar wind parameters. The study reveals that extreme learning machine and its online sequential version are promising models which could predict the AE index extremely fast with a high degree of accuracy even during disturbance periods. The study also shows that the choice of the polar cap index (PC index) as an input parameter is extremely important for an accurate prediction of the AE index.