BP神经网络和ARMA模型在中纬度TEC短期预测中的对比分析

在充分考虑TEC序列非平稳、非线性、高噪声特性前提下, 以IGS提供的2017年电离层TEC格网数据为基准，运用BP神经网络和ARMA两种模型分别进行TEC 3 d预测，重点分析两种模型在不同季节时段、不同电离层活跃强度及不同样本长度下的TEC预测性能及精度。结果表明，在不同时段，两种模型均能很好地反映TEC的变化特性，其中ARMA模型在春、冬时段及整体预测精度上略优于BP神经网络。在平静期，两种模型的平均相对预测精度分别为87.3%和87.5%，预测效果相差较小；在活跃期，两种模型的平均相对预测精度分别为78.5%和75.5%，BP神经网络的精度比ARMA模型高3%。随着样本长度的增加，BP神经网络在21 d样本处预测效果最佳，ARMA模型的预测精度随样本长度的增加呈降低趋势。

Comparative Analysis of BP Neural Network and ARMA Model in Short-Term Prediction of Mid-Latitude TEC

Considering that ionospheric total electron content(TEC)presents as non-stationary,nonlinear and high noise characteristics,we use ionospheric TEC data provided by international GPS service(IGS)to predict 3 days TEC using the back propagation(BP)neural network model and the ARMA model.We compare the prediction performance and accuracy of each model in different seasons,different ionospheric active intensity and different sample lengths.The results show that the two models can well reflect the change characteristics of TEC in different seasons,among which ARMA model is slightly better than BP neural network in spring and winter.In the quiet period,the average relative accuracy of the two models is 87.3%(BP)and 87.5%(ARMA),which means the predictive effect is similar.In the active period,the average relative accuracy of the two models is 78.5%(BP)and 75.5%(ARMA).The accuracy of BP neural network is 3%higher than that of ARMA model.With the increase of sample length,the accuracy of BP neural network model reaches the maximum on the 21st day,the prediction accuracy of ARMA model decreases with the increase of sample length.