Application of combined ICEEMDAN and LSTM in nonlinear coordinate sequence of GNSS
In view of the problem of complex components in the global satellite navigation system(GNSS)nonlinear signals,the difficulty of extracting non structural information effectively and low prediction accuracy of coordinate time series,an improved model combining adaptive noise complete ensemble empirical mode decomposition(ICEEMDAN)and long-term and short-term memory neural network(LSTM)was proposed in this paper.First,ICEEMDAN was used to decompose the GNSS coordinate time series of 10 stations in Japan for 11.4 years.Then the correlation coefficient method was used to divide the decomposed sub series into high-frequency term,low-frequency term and trend term,and the LSTM was used to reconstruct and predict the GNSS coordinate time series.The results showed that:firstly,ICEEMDAN-LSTM method could adaptively extract periodic signals of different frequencies and amplitudes according to the characteristics of different stations.After the correction of periodic terms,compared with the traditional harmonic model,the average weighted root mean square(WRMS)of time series in N,E and U directions were reduced by 8.66％,4.83％and 7.17％,respectively,indicating that the model was more accurate and effective in correcting the original time series in three directions.Secondly,compared with the prediction results of LSTM and EEMD-LSTM models,the average MAE value of the prediction results of this model decreased by 39.31％and 11.19％respectively,and the average RMSE value decreased by 32.00％and 14.34％respectively,indicating that the prediction accuracy of this model was higher,which could be used for GNSS nonlinear coordinate time series prediction.
NETL
NSTL
万方数据
