One day prediction of nighttime VLF amplitudes using nonlinear autoregression and neural network modeling
One day prediction of nighttime VLF amplitudes using nonlinear autoregression and neural network modeling
复制标题
DOI：
10.1002/2016rs006022
复制
发表时间：
2017-01
期刊：
影响因子：
1.6
通讯作者：
H. Santosa;Y. Hobara
中科院分区：
文献类型：
--
作者：
H. Santosa;Y. Hobara
在日本东京乔夫（CHF）连续记录了来自夏威夷（NPM）的甚低频（VLF）发射机的电场幅度。VLF振幅的变化表明NPM-CHF传播路径附近D区（60-90 km高度范围）存在低电离层扰动。NARX神经网络模型，这是建立在各种物理参数，如平流层温度，总柱臭氧，宇宙射线，Dst和Kp指数的日常输入变量的基础上具有良好的精度在建模过程中。在2011年1月1日至2013年2月4日的训练期内，通过使用三种算法，即贝叶斯神经网络（BRANN）、Levenberg马夸特神经网络（LMANN）和标度共轭梯度（SCG），构建了拟合模型。LMANN具有最大的Pearson相关系数（r）0.94和最小的均方根误差（RMSE）1.19 dB。利用LMANN建立的模型对2013年2月5日至2013年12月31日的VLF振幅进行了预测。结果，提前一步（1天）预测的夜间VLF振幅的r为0.93，RMSE为2.25 dB。我们的结论是，根据所提出的方法建立的模型提供了很好的预测的电场振幅的VLF波的NPM-CHF（中纬度）传播路径。
The electric field amplitude of very low frequency (VLF) transmitter from Hawaii (NPM) has been continuously recorded at Chofu (CHF), Tokyo, Japan. The VLF amplitude variability indicates lower ionospheric perturbation in the D region (60–90 km altitude range) around the NPM‐CHF propagation path. We carried out the prediction of daily nighttime mean VLF amplitude by using Nonlinear Autoregressive with Exogenous Input Neural Network (NARX NN). The NARX NN model, which was built based on the daily input variables of various physical parameters such as stratospheric temperature, total column ozone, cosmic rays, Dst, and Kp indices possess good accuracy during the model building. The fitted model was constructed within the training period from 1 January 2011 to 4 February 2013 by using three algorithms, namely, Bayesian Neural Network (BRANN), Levenberg Marquardt Neural Network (LMANN), and Scaled Conjugate Gradient (SCG). The LMANN has the largest Pearson correlation coefficient (r) of 0.94 and smallest root‐mean‐square error (RMSE) of 1.19 dB. The constructed models by using LMANN were applied to predict the VLF amplitude from 5 February 2013 to 31 December 2013. As a result the one step (1 day) ahead predicted nighttime VLF amplitude has the r of 0.93 and RMSE of 2.25 dB. We conclude that the model built according to the proposed methodology provides good predictions of the electric field amplitude of VLF waves for NPM‐CHF (midlatitude) propagation path.