Dynamic mode decomposition and bivariate autoregressive short-term prediction of Earth rotation parameters

Abstract

Abstract In this contribution two new approaches are applied to predict polar motion and length-of-day. The first one is based on Dynamic Mode Decomposition (DMD), that is purely data-driven and is capable of reconstructing and forecasting time series in one numerical procedure. The other one is based on a vector autoregression of order p – VAR(p), which is a vector counterpart of AR(p) that accounts for an evolution of variables in time and a coevolution with other variables. DMD was applied to polar motion and length-of-day whilst VAR(p) to a joint prediction of polar motion. A prediction experiment concerned 30-day forecast horizon with a 7-day shift. It was performed separately for years 2017–2022 giving 48 predictions within each year. This study uses IERS EOP 14 C04 (IAU2000) as a reference for all computations and a mean absolute prediction error (MAPE) as a measure of prediction quality. For DMD, MAPEs for x coordinate of the pole vary from 0.22–0.30 mas for the 1st day and 6.64–8.56 mas for the 30th day of prediction depending on the year whilst those values vary from 0.20–0.27 mas and 5.27–7.66 mas for VAR(p) based prediction. Corresponding values for y coordinate of the pole vary from 0.15–0.23 mas and 4.27–5.93 mas for DMD, whilst 0.13–0.21 mas and 3.46–3.82 mas for VAR(p). In case of LOD forecast, MAPEs vary from 0.023–0.031 ms for the 1st day and 0.142–0.205 ms for the 30th day depending on the year.