The Significance of Input Features for Domain Adaptation of Spacecraft Data

Abstract

The problem of improving the neural network forecast of geomagnetic index Dst under conditions in which the input data for such a forecast are measured by two spacecraft, one of which is close to the end of its life cycle, and the data history of the other is not yet enough to construct a neural network forecast of the required quality. For an efficient transition from the data of one spacecraft to the data of another, it is necessary to use methods of domain adaptation. This paper tests and compares several data translation methods. Also, for each translated attribute, an optimal set of parameters for its translation were found, which further reduces the difference between domains. The paper shows that the use of domain adaptation methods with the selection of significant features can improve the forecast compared to the results of using untranslated data.