Modeling Of SuperDARN Ionospheric Convection Using Deep Learning

Plasma convection is a characteristic phenomenon in the high-latitude ionosphere,and is a manifestation of the coupling of the solar wind and the magnetosphere in the ionosphere.It contains a series of important information about the energy transfer of the solar wind to the magnetosphere,ionospheric system.The research on the evolution of ionospheric plasma convection in time and space is important for the in-depth understanding of the magnetosphere-ionosphere coupling mechanism,and the subsequent modeling and prediction of space weather.On the basis of several deep learning algorithms,this thesis models the data of ionospheric convection with high spatiotemporal resolution which were detected in northern hemisphere by Super DARN radar network.A few prediction models of plasma convection map for high-latitude ionosphere are formulated.In addition,using the datasets collected in independent testing periods,this thesis analyzes and assesses the performance of these models on four indicators such as structural similarity(SSIM),root mean square error(RMSE),mean absolute error(MAE)and linear correlation coefficient(LC)based on the values of radar measurements and models.The results show that the overall performance of spatiotemporal sequence models is appealing,among which ED-Conv LSTM makes the best job as its prediction convection map is the closest to the real image measured by the radar.This study proposes a new idea for the futher research on ionospheric convection as its first modeling and prediction on convective map in this field which is different from previous studies that model and forecast a certain parameter of ionospheric convection by deep learning algorithms.In addition,this thesis also calculates three important convective parameters like cross polar cap potential(CPCP),cross polar electric field(CPEF)and vortex spacing with the help of the constructed model of ionospheric convection graph.The estimation has been done on the results via RMSE,MAE and statistical distribution,too.The simulations demonstrate that all models work well,and the ED-Conv LSTM,a spatiotemporal sequence model,still on the top of the others in a comprehensive comparison.Previous studies often build different models to predict the corresponding parameters,but the model in this thesis is more general and can predict multiple convective parameters at the same time.Finally,this article also discusses the impact of the prediction interval on the performance of the model.According to the time,a prediction model of 2 minutes,10 minutes,30 minutes and 60 minutes is constructed.The results show that as the prediction interval increases,the performance of the model will gradually decline,and the speed of decline of different models is also different.The performance of the spatiotemporal sequence model ED-Conv LSTM with encoder and decoder structure is the best,and the performance degradation rate is the slowest,which shows its reliability in ionospheric convection modeling.In conclusion,this thesis not only proves the effectiveness of deep learning technology in ionospheric convection modeling,but also proves the high accuracy and reliability of spatiotemporal sequence models in various deep learning models.