Aiming at the difficulties in extracting the constructive and non-constructive information of GNSS coordinate time series effectively and identifying and removing noise effectively, this paper pro-poses an improved model combining improved complete ensemble EMD (ICEEMDAN) and long short-term memory (LSTM). Firstly, the GNSS time series was decomposed by ICEEMDAN, and then the decomposed components were divided into high frequency, low frequency and trend items by using Hausdorff Distance (HD). Finally, the signal was reconstructed by LSTM. The results of simulation experiment and real data experiment (Yunnan region) show that the model can effectively extract the nonlinear part of coordinate time series. Compared with continuous root mean square error (CMSE), energy density (ET) and correlation coefficient rule, HD can effectively identify and eliminate the noisy part of the signal. In addition, compared with the linear motion model, this model can effectively extract the nonlinear trend part of the signal. After the correction of period terms, the average RMS of the model in the E, N and U directions decreased by 46.57%, 35.09%, 44.48% and 15.11%, 4.10% and 3.42%, respectively, compared with that of the harmonic model and the ensemble EMD (EEMD), indicating that the model can extract the period information more fully. In addition, this paper also proposes to combine the model with the environmental load to weaken the nonlinear changes in the signal. The results show that: The mean RMS decreases by 14.32%, 9.46% and 4.05%, respectively, when hydrographic load, non-tidal atmospheric load and Marine load are added in the U direction. There is no significant change in the horizontal RMS, indicating that the elevation direction of the station in Yunnan region is susceptible to environmental load. It is also found that the smaller the phase difference between the load series and the load series after load correction, the closer the annual cycle amplitude, the better the effect of load correction.
