利用神经网络预报电离层f_0F_2

Forecasting of Ionospheric Critical Frequency Using Neural Networks

由中国武汉电离层台站和澳大利亚Hobart台站的电离层F2层临界频率(f0F2)的资料,利用三层 前向反馈神经网络(BP网络),提出一种提前24 h预测,f0F2的方法,该方法以前5天观测的,f0F2数据拟合 的5个系数以及太阳活动参数作为输入,以当天24 h的,f0F2作为输出对网络进行训练,训练好的网络可以实 现对,f0F2提前24 h的预报.预测结果显示,利用神经网络预测的,f0F2与实际观测结果变化趋势较一致,并且 比IRI的计算结果更加准确.误差分析表明,在南半球Hobart(-42.9°,147.3°)台站比中国武汉站(30.4°, 114.3°)的结果要好,在低年比高年要好,在冬夏季节比春秋季节稍好.本文说明利用神经网络对电离层参量进 行预报是一种切实可行的方法.

The use of feed-forward back propagation neural networks to predict ionospheric F2 layer critical frequency, f0F2, 24 h ahead, have been examined. The data we used are from Wuhan ionospheric station, China, and Hobart ionospheric station, Australia. The data period is from 1970 to 1990 at Wuhan and from 1962 to 1990 at Hobart. The five day's measurements of f0F2 before the day that need forecast are reduced to five coefficients. The inputs used for the BP neural network are the coefficients, the solar 10.7cm flux index, and the outputs are the day's 24 h observed f0F2 data. The trained net then can forecast f0F2 24 h advance. The result indicates the predicted f0F2 using NN has good agreement with observed data. Comparison with IRI model suggests that NN method is more accurate than IRI. In addition, the error analysis indicates that predicted f0F2's Root-Mean-Square Error (RMSE) is smaller in Hobart than in Wuhan, smaller in low solar activity than in high solar activity, smaller in winter and summer than in spring and autumn. In conclusion, using neural network to predict ionospheric parameters is a feasible method.