Confirmation of the Effect of Simultaneous Time Series Prediction with Multiple Horizons at the Example of Electron Daily Fluence in Near-Earth Space

摘要

It is often necessary to make time series (TS) predictions for several values of the prediction horizon. Usually such predictions are made in autonomous mode, i.e. separately for each horizon value. Meanwhile, it is also possible to make simultaneous predictions for all the desired horizons, or group prediction for several horizons at once. In the preceding studies [1], it has been demonstrated that group determination of parameters in solving multi-parameter inverse problem with a multi-layer perceptron (MLP) may outperform autonomous determination if the approximated dependences of the grouped parameters on the input features of the problem are similar and if the sets of significant input features largely intersect. Last year it has been demonstrated, that the effect also holds for MLP TS prediction with multiple horizons [2]. In the present study, efficiency of group prediction of TS with MLP has been checked at the example of TS of electron daily fluence in near-Earth space, which is characterized by rapid degradation of prediction quality with increasing horizon. Relativistic electrons (RE) of the outer Earth's radiation belt are sometimes called "killer electrons" since they can damage electronic components, resulting in temporary or even complete loss of spacecraft. Daily fluence is summary daily flux of these electrons; at geosynchronous orbit of about 35,000 km altitude it is of interest due to the large number of satellites populating this region, and it is predictable thanks to long TS of experimental data available. For this problem, group prediction with average size of groups proved to outperform autonomous and simultaneous prediction. Thus, the positive effect of group determination of outputs in multi-output problem has been confirmed as a property of MLP as data processing algorithm. This study has been performed at the expense of Russian Science Foundation, project no. 16-17-00098.