Research On Determination Of Drag Coefficient Of Low RSO Based On Random Forest

Low-Earth-Orbiting(LEO)satellites play an important role in Earth observations and science missions.As more and more satellites are put into orbits and the number of space debris grows,the collision risk between LEO objects has significantly increased.In order to avoid collisions effectively,it is necessary to have the precise orbit determination(OD)and orbit prediction(OP)capabilities.The accurate perturbing force models,as well as high-precision and dense observations,are the foundation for the precise OD.However,for most LEO objects,especially space debris,tracking data is usually both sparse and low accurate.In addition,the accuracy of atmospheric mass density models(ADMs)is low.As such,the 7-day orbit prediction error of LEO debris could reach tens or even hundreds of kilometers.A main cause is the large errors in the drag coefficient,,which calls for the study on the high-precision drag models for LEO objects.The drag coefficient is a critical parameter for the orbit determination and prediction of a LEO object.Experiments have shown that,under the sparse data condition,the drag coefficient estimated in the orbit determination process is not suitable for the orbit prediction.On the other hand,the drag coefficient has complicated temporary-spatial variation property,and it is still a challenging task to model it.With the objective of OP accuracy improvement,this thesis proposes to model the drag coefficient from the use of historical data with Random Forest Regressor-based approach.The main research work and innovations are as follows.1.Research on three methods of computing the drag coefficient is performed.Through simulation experiments considering different scenarios of observations,ADMs,and space environments,the sensitivity of the OP errors to the drag coefficient is analyzed.2.Drag coefficient modeling with the Random Forest Regressor is introduced.The basic principle of decision tree and the integration idea are discussed,and the basic process applying the Random Forest is presented.The algorithm of the drag coefficient prediction model based on the Random Forest is designed via Python.The important parameters in the model construction are studied by experiments.3.A Random Forest model of predicting the drag coefficient of the GRACE-A satellite using historical drag coefficients is constructed and the performance is assessed.The assessment results show that,in the high solar activity time,the drag coefficient predicted from the Random Forest model results in much smaller 7-day OP errors than the use of the fitting coefficient.4.The sparse radar tracking data scenario is simulated for the GRACE-A and CHAMP satellites.A grid-search method is proposed to determine drag coefficient series from sparse tracking data.The determined drag coefficient series is then used to construct the Random Forest prediction model.The 7-day OP errors for both the GRAC-A and CHAMP satellites are significant reduced when the Random Forest prediction models are used.The study on the drag coefficient prediction presents a promising method to improve the OD/OP accuracy for LEO debris.The thesis also has developed a software platform to make experiments of using real tracking data and testing other machine learning algorisms.