Performance Analysis of Distinct Feed-forward Neural Networks Structures on the AE Index Prediction

Abstract

Space weather events, such as geomagnetic storms and substorms, can affect the operation of modern technological systems that mankind currently relies on, including electric power grids, spaceborne vehicles, radio wave communication and navigation systems. Therefore, any new information that helps us to have a better understanding of the space weather and their impact on our infrastructure and services is of the highest interest. Furthermore, monitoring and forecasting these disturbances are essential activities to avoid damage and losses. Over the past decades several different indices have been developed in order to monitor the space weather. One index commonly used by the scientific community to assess the level of geomagnetic activity in the auroral region is the AE (auroral electrojet) index, which reflects changes on the auroral electrojets due to the fluctuations in the solar wind convection electric field that are produced by variations in the solar wind velocity and interplanetary magnetic field (IMF). In this framework, this paper investigates the performance of distinct arrangements of feedforward artificial neural networks (ANNs) on the estimation of the hourly-averaged AE index aiming at monitoring and forecasting activities. We have analyzed the effect of the inclusion of different information related to the IMF, solar wind and their daily and seasonal variability using more than 10 years of data derived from the Advanced Composition Explorer (ACE) and Wind missions. Our analysis indicates that the inclusion of the temporal information improved the performance of the ANN. Tests on the most suitable arrangement during different periods during the year of 2012 revealed a good correlation between the estimated and the reference AE index values and indicate that the model could reproduce the component of the AE index that is directly driven by the solar wind. The ANN model was able to qualitatively depict long temporal variations of the hourly averaged AE index during periods of low, moderate and high geomagnetic activity.