Application Research Of Deep Learning Method Of Self-generating Network In 4D Track Prediction

Civil aircraft flying in accordance with 4D track is carrying out fine control of spatial three dimensions and time dimension in all stages of take-off,climb,cruise,descent and landing,and the operation based on 4D track will be the core operation concept of the future air traffic control system.The air traffic situation is complex and changeable,and the trajectory of the aircraft will be affected by many factors,like weather and control.Nevertheless,how to make full use of the position and status information of a flight in the real-time system to analyze and predict rapidly and effectively,providing substantial reference information for the system or controller,has been a research hotspot in this field.With the promotion of broadcast automatic dependent surveillance(ADS-B),the real-time position and dynamics of each aircraft can be acquired by ground stations and other aircrafts.High-frequency historical trajectory data and real-time data that provides valuable information can be used for pattern mining and pattern recognition in aircraft trajectories.However,the current problems are: the trajectory points obtained by iterative 4D trajectory prediction are relatively sparse,with the real-time performance not strong;the utilization rate of ADS-B data is low,with the information not fully mined;the method of automatically constructing deep neural network application models has a high cost of time and a large demand for hardware.Therefore,in combination with the needs of air traffic management(ATM),this paper deeply analyzes a large amount of ADS-B historical data,and studies the deep generation models,feature extraction algorithms and network structure self-generation algorithms that can do prediction directly.For the specific task of 4D track prediction,a step-by-step model structure selfgeneration strategy and a 4D track prediction algorithm based on a self-generated deep neural network are proposed,and the model generation process and the algorithm’s prediction accuracy,time delay,etc.are analyzed through experiments,verified the effectiveness of the strategy and the algorithm.In addition,this paper also proposes a data compression strategy for high-density ADS-B data,which improves the time efficiency of model structure search.This article decomposes the research content from the perspective of system identification,from data analysis and preprocessing,to theory analysis and technic selection,to the proposal and verification of strategies and algorithms,and finally through a 4D track-based traffic prediction application scenario,the accuracy and efficiency of the model are verified.Experimental results show that compared with other research and implementation methods,the strategy proposed in this paper can generate the required application model more efficiently and is scalable;the proposed algorithm realizes the 4D track prediction task with higher prediction accuracy and stronger real-time performance.