Empirical Model Of Ionospheric Total Electron Content Using Machine Learning Technique

Ionosphere is an important part of terrestrial space.Total Electron Content(TEC)is proportional to the time-delay error of satellite signals' passage through ionosphere.Because the TEC deviate from the quiet value seriously when there are interplanetary disturbances passing the Earth,researching TEC storms' physical process and forecasting the TEC have wide spread application value.This thesis summarized the ionospheric storm response to interplanetary solar wind and magnetosphere using CODE TEC map and developed an empirical model on forecasting storm-time TEC by deep learning recurrent neural network.The main contents and results are presented as follows:1.The response of TEC in 110°E of China to magnetic storm on March 17-18,2015 was analyzed.Solar wind parameters,magnetic index and polar cap potential are used to discussed the physical process of TEC storm.Results showed that there was an initial positive storm in high-latitude and mid-latitude during the main phase on March 17.There was a strong negative storm from high to low latitude on March 18.The space environment indicated that the IMF Bz was southward for a long time,during which solar wind energy entered into the polar thermosphere,with ionosphere lifted,N2/O transferred to lower latitude and dynamo electric field disturbed.All above leaded to the initial positive storm in high-mid latitude and negative storms from high to low latitudes.Besides,there was no strong upward drift in low-latitude ionospheric F region,which means there was no prompt penetration electric field and it may be the reason for lack of positive storm in low-latitude.2.A deep learning recurrent neural network was developed for forecasting ionospheric total electron content at Beijing station 24 h ahead.The predictors included solar 10.7 cm flux index,geomagnetic ap index,grid map of TEC,solar wind speed and the southward components of interplanetary magnetic field.The validation showed that the root-mean-square error of quiet ionosphere TEC predicted by RNN model is about 3.0 TECU and less than the RMSE forecasted by BP model.Positive storm predictability of models with solar wind parameter is increased by 7.5% from other models.Furthermore,the hit rate of the RNN model was 29% higher than the result of BP network,whereas the RMSE of TEC was 0.3 TECU smaller than that of BP model.The RMSE on 31 TEC storms with solar wind parameters included in RNN model decreased 0.45~0.65 TECU.Our comparison indicated that RNN model for short-term forecasting of TEC is more reliable than BP.Moreover,interplanetary solar wind parameters are effective in predicting TEC positive storm.3.A regional model of ionospheric TEC in China was developed.The input data concluded solar wind parameters,magnetic index,solar activity index and TEC data for a regional grid from 35°N ~45°N and 100°E ~120°E.Another regional TEC forecast model was developed for comparison with a multi-point TEC input data.The results showed that the root-mean-sqrt error of grid regional model was less than the multi-point regional model because it could catch the transport process along different latitude,and the TEC in Beijing forecasted from the regional model was even better than that from single station model.However,the prediction accuracy of TEC in high-mid latitude is better than that of low latitude due to the equatorial ionization anomaly.