Satellite Orbit Prediction Method Based On Compensation Model By Using Deep Neural Network

Satellite orbit prediction is the foundation of satellite application research. In order to improvethe accuracy of satellite navigation and positioning, the high accuracy model of orbit predictionshould be formulated. High accuracy prediction model is also very important when it comes to spacemission like satellite autonomous navigation and spacecraft rendezvous and docking. This paperemphasizes on the method which can improve the accuracy of navigation satellite orbit prediction, themain contents and innovations are summarized as follows:1. The dynamic model of navigational satellite with meter level accuracy is introduced in detail,various perturbative forces and their influence on navigational satellite are analyzed. A dynamicalmethod to predict the orbit of GPS navigational satellite is presented. Short term and long termprediction simulation are made and simulation result is analyzed.2. The topological structure and training algorithm of deep neural network are introduced. Aschematic diagram of training algorithm is given. Restricted boltzmann machine is the cornerstone ofdeep neural network, the core of its training algorithm is sampling from the model. There Markovchain Monte Carlo sampling algorithms and structural characteristics of Restricted boltzmannmachine are analyzed, blocked Gibbs sampling algorithm which fits Restricted boltzmann machinebest is adopted in simulation.3. Dynamic model’s orbit prediction error is analyzed. A hybrid orbit prediction model based ondeep neural network is established, contrast experiment are done with GPS navigational satellites.Contrast experiment result show that hybrid model based on BP network mainly compensatedivergence term, however, hybrid model based on deep neural network not only compensatedivergence term but also compensate periodic term in a large degree.