利用2017年中低纬电离层总电子含量、地磁活动指数、年积日等参数，首次建立基于贝叶斯正则化(Bayesian regularization)的Elman回归神经网络(BR-Elman)的电离层TEC预报模型。同时，根据地磁活动指数的变化特征，分别进行平静电离层和扰动电离层预报建模。实验结果表明，该方法在平静期5 d预测值的均方根误差为1.19 TECu，残差为1.03 TECu，相关系数为0.93；在扰动期5 d预测值均方根误差为1.34 TECu，残差为1.01 TECu，相关系数为0.91。贝叶斯正则化的BP神经网络模型以及传统BP神经网络模型在平静期与扰动期5 d的预测上，均方根误差最小为1.87 TECu，残差最小为1.50 TECu，相关系数最优为0.87。通过对比分析，该模型较其他2个模型的预报效果有明显改善。"/>
We use the ionospheric total electric content(TEC), geomagnetic activity index, day of year and other parameters of mid-and low-latitude in 2017 to establish an ionospheric TEC prediction model based on Elman neural network with the Bayesian regularization(BR-Elman). According to the variation characteristics of the geomagnetic activity index, we construct the ionosphere model under quiescent ionospheric conditions and disturbed ionospheric conditions, respectively. The experimental results show that root mean square error, residual error and correlation coefficient of the predicted value of 5 d using the proposed method in the quiet period are 1.19 TECu, 1.03 TECu, and 0.93, respectively. The root mean square error, the residual error and the correlation coefficient of predicted value of 5 d using the proposed method in the disturbed period are 1.34 TECu, 1.01 TECu, and 0.91, respectively. The minimum root mean square error is 1.87 TECu, the minimum residual is 1.50 TECu, and the optimal correlation coefficient is 0.87 by the BP neural network model with Bayesian regularization(BR-BP) and the traditional BP neural network model. The results show that predicted accuracy of the proposed method has improved significantly compared with the prediction effects of the other two models."/>
贝叶斯正则化的Elman神经网络电离层TEC预报模型
Abstract：We use the ionospheric total electric content(TEC), geomagnetic activity index, day of year and other parameters of mid-and low-latitude in 2017 to establish an ionospheric TEC prediction model based on Elman neural network with the Bayesian regularization(BR-Elman). According to the variation characteristics of the geomagnetic activity index, we construct the ionosphere model under quiescent ionospheric conditions and disturbed ionospheric conditions, respectively. The experimental results show that root mean square error, residual error and correlation coefficient of the predicted value of 5 d using the proposed method in the quiet period are 1.19 TECu, 1.03 TECu, and 0.93, respectively. The root mean square error, the residual error and the correlation coefficient of predicted value of 5 d using the proposed method in the disturbed period are 1.34 TECu, 1.01 TECu, and 0.91, respectively. The minimum root mean square error is 1.87 TECu, the minimum residual is 1.50 TECu, and the optimal correlation coefficient is 0.87 by the BP neural network model with Bayesian regularization(BR-BP) and the traditional BP neural network model. The results show that predicted accuracy of the proposed method has improved significantly compared with the prediction effects of the other two models.
