GNSS坐标非线性变化的差分长短时记忆网络预测

Prediction of GNSS coordinate nonlinear variations using difference method and long short term memory

为了改进GNSS坐标非线性变化的预测方法,获得更高的预测精度,该文提出了一种基于一阶差分和长短时记忆(LSTM)网络的坐标非线性变化预测方法。首先综合利用奇异值分解和最大互信息系数准则对坐标时间序列进行降噪处理得到真实的非线性变化,接着利用谐波模型对其中的周期项进行提取和预测,剩余的未模型化成分经过一阶差分后采用LSTM网络进行预测,然后将两个预测结果进行综合得到坐标非线性变化的高精度预测结果。实验结果显示,该方法在20 d的预测步长内的平均绝对误差达1 mm以内,相比谐波模型、ARIMA模型和未经一阶差分的LSTM模型的预测方法精度至少提升了78%、25%和22%,具有更高的预测精度。同时经过对比也证明了该方法具有更好的适用性。

In order to improve the prediction method of GNSS coordinate nonlinear variations and obtain higher prediction accuracy,a method based on first-order difference and long short term memory(LSTM)was proposed.First,the singular value decomposition and maximal information coefficient criteria were used to denoise the coordinate time series to obtain the real nonlinear variations.Then the harmonic model was used to extract and predict the period components,and the remaining unmodeled component was predicted by the LSTM network after first-order differenced.Finally,the two prediction results were synthesized to obtain a high-precision prediction result of coordinate nonlinear variations.The experimental results showed that the method could obtain a prediction result with its mean absolute error less than 1 mm within the prediction step of 20 d,improving by at least 78%,25%and 22%compared to the method using the harmonic model,autoregressive integrated moving average(ARIMA)model and undifferenced LSTM.It indicated that the method had higher prediction accuracy.Meanwhile,the result also proved that this method had better applicability.