【摘要】 在GPS时间序列数据补全问题中,针对在空间点位分布稀疏、观测值连续缺失情况下传统补全效果不佳的问题,提出了一种基于深度学习的补全方法。针对时间序列中存在连续大量缺失的情况,设计了基于长短时记忆神经网络的补全模型;使用待补全站时间序列中可用的数据训练模型,合并使用日期数据增强训练效果,使模型学习到隐式蕴含在序列中的时空相关知识,预测序列缺失处的值。用IGS基准站SHAO的1999—2017年有缺失的时间序列进行实验;并与时间序列预测的传统经典方法 Seasonal ARIMA进行了比较。实验结果表明:在待补全站观测值缺失较多时,提出的方法依然可以取得很好的补全效果;在补全较长连续缺失时,无论是在预测精度还是对原始序列形态的模拟上,表现均优于Seasonal ARIMA。更多还原

【Abstract】 A novel method of reconstruction of gappy GPS time series based on the deep learning approach is proposed to conquer the problem that traditional ways of interpolating data can’t perform well while spatial monitoring points are sparse and data missing gaps are wide. A Long-Short-Term Memory( LSTM) based on deep neural network is built to learn the spatial and temporal knowledge implicitly embedded in GPS time series. For a gappy time series to reconstruct, its available historical data is used to train the network, then the values of missing points are predicted by the network for reconstruction. Time stamp data and some statistical values of the time series are merged into the learning process to enhance the performance. Daily coordinate time series of IGS site( SHAO, from1999 to 2017) is used to validate the effectiveness of the method. Experiment using Seasonal ARIMA is performed for comparison. Results show that the proposed method achieves satisfactory results when using only limited available data of a severe gappy time series. It outperforms Seasonal ARIMA, especially for situations with wide gaps.The proposed method not only achieves smaller RMSE on the point predications, but also the time series segments it produced simulate the shape of the targeting time series, comparing to the Seasonal ARIMA method.更多还原