Research And Application Of GRU Neural Network In Trajectory Prediction

The rapid development of the air transportation industry has brought much convenience to life,but the tension of airspace resources has been aggravated.At the same time,the pressure of air traffic controllers has been increased sharply.Accurate four-dimensional trajectory prediction technology is the key to solve the shortage of airspace resources and reduce the pressure of air traffic controllers.Moreover,an air traffic control automation system based on4 D trajectory prediction can assist controllers to ensure flight safety,increase airport throughput and reduce airline operating costs.Firstly,this thesis elaborates on the domestic and foreign research status of 4D trajectory prediction technology.A short-term 4D trajectory prediction algorithm based on onlineupdating GRU neural network can be used to solve the problem that some factors of the flight process cause the real-time trajectory to have a certain difference compared with the historical trajectory to make the big error of prediction.First,the historical trajectory prediction model is established based on the GRU neural network and trained by using the historical flight data.The parameters of the historical prediction model have been trained are saved.When receiving the flight trajectory data,the received real-time trajectory data is used to retrain and fine-tune the parameters in the historical prediction model.The retrained and fine-tuned historical prediction model is used as the online-updating prediction model.The trajectory points and arrival time in the future are predicted.Subsequently,further research found that the aircraft has an obvious flight phase in a complete flight mission.Therefore,a short-term 4D trajectory prediction algorithm combining K-means clustering and online-updating GRU is proposed.Since the aircraft exhibits unique motion trajectory characteristics in each flight stage,the trajectory points with similar motion characteristics are clustered into the same cluster.Then,the online-updating GRU short-term4 D trajectory prediction model is used to analyze each cluster separately which makes it easier for the neural network to learn this kind of motion characteristics.Finally,predictions are made about the trajectory points that will be reached in the future.The real trajectory data,which are the fusion data by the secondary radar data and the automatic related surveillance-broadcast data,is used to verify the two algorithms given in this thesis.The experimental results show that the two prediction algorithms have better prediction performance.