Application of the radial basis function neural network to the short term prediction of the Earth’s polar motion

Abstract

By a number of test cases using different sample numbers and sample lengths, we obtain a Radial Basis Function Neural Network (RBFNN) model that is suitable for the short-term forecast of polar motion, especially for the ultra-short-term forecast. By using the same data sample of Earth’s polar motion, this RBFNN model can achieve better short-term prediction accuracy than the least-squares+autoregressive (LS+AR) method, and better ultra-short-term prediction accuracy than the LS+AR+Kalman method. Using this model to forecast the polar motion data from January 1, 2002 to December 30, 2007 and from January 1, 2010 to December 30, 2016, respectively, experimental results show that the ultra-short-term forecast accuracy of this RBFNN model is within a precision of 3.15 and 3.08 milliseconds of arc (mas) in polar motion x direction, 2.02 and 2.04 mas in polar motion y direction; the short-term forecast accuracy of RBFNN model is within a precision of 8.83 and 8.69 mas in polar motion x direction, and 5.59 and 5.85 mas in polar motion y direction. As is stated above, this RBFNN model is well capable of forecasting the short-term of polar motion, especially the ultra-short-term.