Modeling Of SuperDARN Polar Ionospheric Electric Field Using Deep Learning

Ionospheric convection is an important phenomenon in space weather which varies with the changes of the interplanetary magnetic field(IMF)and solar wind parameters,and is also closely related to substorms.The research on convective electric field and convection model in ionosphere has been a hot topic for many years.In this paper,the polar ionospheric convective electric field is calculated by Cross Polar Cap Potential(CPCP)of the Super DARN.Based on the data of the polar ionosphere convective electric field in 2014,12 parameters relate to the convective electric field are selected according to previous studies,and the historical data with the time delay of 20minutes of convective electric field are introduced.The data are preprocessed,the invalid data are deleted and the data with echo points greater than 300 are selected.After processing the data,we find that the 13 radar parameters are too complex for the construction of ionospheric electric field model,so we need to reduce the dimension of them.The feature extraction is based on the correlation between parameter data and convective electric field;Principal component analysis(PCA)and singular value decomposition(SVD)are used for feature extraction.Then,the ionospheric electric field models are constructed based on back propagation neural network(BPNN),long-short term memory neural network(LSTM),autoregressive(AR)and multiple linear regression(MLR)algorithms respectively.The performance of the models bases on dimension reduction matrix obtained by different data dimension reduction methods is studied.The evaluation parameters include the distribution of predicted values of the eletric field model,absolute deviation,relative deviation,mean absolute error(MAE),root mean square error(RMSE)and linear correlation coefficient(LC).The results show that:the performance of LSTM model bases on SVD analysis after correlation feature selection is the best.The absolute error and relative error are the smallest.And RMSE=2.83m V·m^-1,MAE=2.14m V·m^-1,LC=0.80,all of them achieve optimal results.The performance of BP model bases on PCA analysis after correlation feature selection is the second.But it also has excellent performance.The absolute error and relative error are small.And RMSE=2.90m V·m^-1,MAE=2.23m V·m^-1,LC=0.80.The conclusion shows that the deep learning algorithm is more effective and accurate than the physical model in nonlinear modeling and prediction of convective electric field.