基于多源数据的GNSS高程预测

GNSS Elevation Prediction Based on Multi-Source Data

提出一种改进的自适应噪声完全集合经验模态分解(improved complete ensemble empirical mode decomposition with adaptive noise,ICEEMDAN)、核主成分分析(kernel principal component analysis,KPCA)和长短期记忆网络(long short-term memory,LSTM)的组合预测模型,以提高GNSS高程时间序列的预测精度。考虑大气压力、地表温度和水文负载对站点位移的影响,首先使用ICEEMDAN对多源环境序列进行分解,降低其非平稳性;然后使用KPCA提取关键影响因子,减少数据冗余与相关性,降低输入维度;最后利用LSTM模型对多维时间序列集GNSS高程时间序列进行建模,实现对GNSS高程的预测。对4个GNSS站点6000组历元数据进行测试的结果显示,ICEEMDAN-KPCA-LSTM模型在预测GNSS高程时间序列时,RMSE和MAE分别比其他模型降低32.8%~35.3%和34.2%~42.1%,皮尔森相关系数提高11.6%~14.3%,表明该组合预测模型具有更好的预测性能。

To enhance the prediction accuracy of GNSS elevation time series,we propose a hybrid forecasting model combining improved complete ensemble empirical mode decomposition with adaptive noise(ICEEMDAN),kernel principal component analysis(KPCA),and long short-term memory(LSTM).Considering the impact of atmospheric pressure,surface temperature,and hydrological load on station displacement,we reduce the non-stationarity of the multi-source environmental sequences by decomposing them using ICEEMDAN;then,we extract key influencing factors using KPCA to reduce data redundancy and correlation,thereby reducing the input dimension;finally,we use the LSTM model to model the multi-dimensional time series set of GNSS elevation time series for prediction.Test results on 6000 sets of historical data from 4 GNSS stations show that the ICEEMDAN-KPCA-LSTM model reduces the RMSE and MAE by 32.8%to 35.3%and 34.2%to 42.1%,respectively,compared to other models,and improves the Pearson correlation coefficient by 11.6%to 14.3%,indicating that this hybrid forecasting model has better predictive performance.