深度学习LSTM模型的电离层总电子含量预报

TEC prediction of ionosphere based on deep learning LSTM model

针对TEC时间序列高噪声、非线性和非平稳的动态序列的特性,基于分解-预测-重构的思想,运用总体经验模态分解和深度学习长短期记忆神经网络,构建了EEMD-LSTM预测模型。同时,以测试集上预测结果的均方根误差最小为目标,运用多层网格搜索算法对EMD-LSTM预测模型进行参数优选。以IGS中心2015年全年1 h时间尺度的TEC格网数据进行实验分析,结果表明,EEMD-LSTM组合模型的预报结果能够很好的反应电离层TEC的变化特性,在低、中、高纬度地区平均预报残差分别为1.37、0.82和0.96个TECu,预测平均相对精度分别为92.8%、91.9%和87.8%。

The total electron content(TEC)is a representative parameter.For its non-linear and non-stationary characteristics taking the TEC data which varies from high latitude to low latitude in both quiet and active period provided by the IGS as sample data.A new combined forecasting model is built in this paper by using ensemble empirical mode decomposition(EEMD)and Long-Short Term Memory Model(LSTM).Furthermore,a multilayer grid search algorithm is proposed to optimize the parameters of EMD-LSTM prediction model.Results from numerical experiments show that the predicted results are highly fitted to the actual observation data.The average prediction residuals in different latitudes are 1.37TECu,0.82TECu and 0.96TECu,respectively.The predicted average relative accuracy is 92.8%,91.9%and 87.8%,respectively.