Forecasting Of Ionospheric TEC Using Long Short-Term Memory Network

The ionosphere is the ionized zone produced by the earth’s upper atmosphere,which is affected by solar radiation and cosmic rays.The range of the ionosphere is about 60 km to 1000 km,and the partial ionization or full ionization of the atmosphere produces free electrons and ions,which are different from the neutral atmosphere and can affect the transmission of radio.Total Electron Content(TEC)is the total electron density of the ionosphere.Ionospheric TEC is an important parameter the ionosphere which reveal the changes of ionosphere.Thus,it is meaningful to focus on the research and forecast of ionospheric TEC.The disturbance of ionosphere is closely related to solar activity and geomagnetic activity.In this thesis,ionosphere TEC,solar activity F10.7 and geomagnetic activity index ap are used to forecast TEC of next 24 hours,by using Long Short-term Memory Network(LSTM)neural network.The main contributions are as follows:(1)We proposed the LSTM networks to predict ionospheric TEC.Experimental results are compared with the traditional BP neural network(MLP).The experiment proves that the LSTM network is faster and more stable than the traditional BP neural network in training process.In addition,the Root Mean Square(RMS)error of the LSTM network prediction is 3.47,and the prediction accuracy is much higher than 5.02 of BP neural network.Finally,we fit curves of forecasted TEC values and true value respectively in three cases which are quiet period,disturbance period and period with sudden disturbance.Experiments show that the curve of LSTM is more accurate in the quiet period.When the ionosphere disturbance happened,prediction results of LSTM completely accords with the cycle of TEC change trend,the traditional BP network in such a complex case would predict the totally opposite trend of the real results,it shows that LSTM network can learn the long-term dependence from raw data,while the traditional BP network can’t do that.However,even the LSTM network could not give a relatively accurate forecast when the ionosphere suddenly disturbed in the calm period.Therefore,we further apply Bidirectional Long Short-term Memory Network(Bidirectional LSTM)to improve this problem.(2)We proposed Bidirectional LSTM model to predict TEC in 24 hours,which improved the problem that the prediction error would become larger when the LSTM was suddenly disturbed in the quiet period.In experimental analysis section,we compare RMS error of Bidirectional LSTM,two-layer LSTM network,single-layer LSTM and two-layer MLP network(BP network)separately,RMS error of bidirectional LSTM is 3.34.Emphatically analyzes the bidirectional LSTM sudden disturbance in the quiet period occurred at the time of the forecast results.Experimental results show that bidirectional LSTM can fully exploit the data information of past and future time and improve the prediction performance based on LSTM network.