Auroral Ultraviolet Images for Prediction of the Auroral Electrojet Index

Abstract Solar wind parameters can effectively predict the component of the auroral current system directly driven by the solar wind, but cannot explain the dense westward electrojet formed through the unloading process of the magnetotail. However, auroral ultraviolet images (UVIs) can spatially map the entire variation process of auroral electrojets (AEs). In this paper, auroral UVIs are used for the prediction of AE index for the first time, and a grid feature extraction method based on correlation coefficient selection is proposed for the spatial mapping relationship between the latitude and longitude distribution characteristics of auroral power (AP) and the AE index. In terms of the prediction algorithm, we use the extreme learning machine (ELM) network, which has strong generalization ability, and compare it with the generalized regression neural network (GRNN) and fully convolutional network (FCN). The experimental results show that the method of predicting AE index by auroral UVI image is feasible, and the root mean square error of the prediction results exceeds the expected accuracy, reaching 8.97%. The study also shows that the grid feature extraction method greatly improves the accuracy of ELM network in predicting AE index, and it is also applicable to other prediction networks. The OS_ELM strategy can further reduce the prediction error by about 1.5%, and it tends to saturate with the increase of input data volume.