联合ICEEMDAN和LSTM在GNSS非线性坐标时序中的应用

Application of combined ICEEMDAN and LSTM in nonlinear coordinate sequence of GNSS

针对全球导航卫星系统(GNSS)非线性信号中成分复杂,非构造信息难以有效提取和坐标时间序列预测精度不够的问题,该文提出了一种改进的自适应噪声完全集合经验模态分解(ICEEMDAN)和长短期记忆神经网络(LSTM)结合的模型。先利用ICEEMDAN对日本地区10个测站连续11.4 a的GNSS坐标时间序列进行分解,再使用相关系数法将分解的若干子序列划分为高频项、低频项和趋势项,接着采用LSTM对GNSS坐标时序进行重构和预测。结果显示:ICEEMDAN-LSTM法可以根据不同测站的特性,自适应提取不同频率和振幅的周期信号,且经周期项改正后,与传统谐波模型相比,N、E、U方向时间序列的平均加权均方根(WRMS)分别降低了8.66%、4.83%和7.17%,表明该模型对原始时间序列在3个方向的改正均更加精确有效。其次,与LSTM和EEMD-LSTM模型预测结果的相比,该模型预测结果的平均MAE值分别降低了39.31%和11.19%,平均RMSE值分别降低了32.00%和14.34%,表明了该模型的预测精度更高,可以用于GNSS非线性坐标时序预测。

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.