Improving the Performance of the LS+NN Model for UT1-UTC Forecast with the Edge Extension

The prediction accuracy of UT1-UTC can be easily affected by the edge distortion of least-squares (LS) fitting time-series, referred to as the edge effect in data-processing fields, when periodic oscillations and residuals are separated by LS fitting. In order to alleviate the edge effect, the original UT1-UTC time-series is first extended on both boundaries using the GM (1, 1) grey model in this work. A LS extrapolation model is then set up using the extended time-series to remove the edge distortion to the boundaries. Finally, UT1-UTC predictions are generated by the combination of the edge effect correlated LS (ECLS) extrapolation model and extreme learning machine neural network (ECLS+NN). The numerical experiments show that the edge effect can be remarkably alleviated with the presented approach. In addition, the accuracy of the UT1-UTC short-term predictions by the ECLS+NN method is better than that obtained by the LS+NN approach, but only slighter. The medium- and long-term predictions, however, are noticeably more accurate than those by the LS+NN solution.