Research On Track Prediction Method Based On Trajectory Clustering And LSTM

In recent years,the civil aviation industry has entered a period of rapid development,and the future air traffic will be more and more intensive.Air traffic management is becoming an increasingly important and complex research field.4D trajectory prediction is the core element of the air transportation system,which aims to improve the operation capability and predictabil-ity of air traffic.4D trajectory prediction is the core element of air transportation system,which aims to improve the operation ability and predictability of air traffic.Accurate trajectory predic-tion can effectively solve the problem of airspace resource shortage,and is of great significance in the fields of conflict detection and resolution,cooperative control and so on.With the rapid development of data mining and artificial intelligence technology,more and more historical track data can be used for airspace situation analysis and monitoring,and its effective applica-tion in intelligent air traffic control system has become an inevitable trend.Currently,there are two main methods being used for 4D trajectory prediction: traditional prediction methods based on physical models and machine learning methods based on a sin-gle network.The former is mainly an aerodynamic model,which lacks consideration of the uncertainty of the external environment such as airspace posture? the latter constructs a single network model,and largely ignores the time series characteristics of the track data itself.There-fore,this paper proposes a trajectory prediction model based on trajectory clustering and LSTM.The specific work and contributions are as follows:Firstly,the initial historical track data sample set is obtained by collecting,cleaning and analyzing ADS-B data.Among them,the crawler technology is used to obtain historical track data,and the parsing of radar message data is completed according to the ASTERIX CAT 062 standard,and the track data containing missing field values,duplicate records and abnormal values in the track are filtered and rejected.Through visual analysis of all airport distributions as well as latitude,longitude,altitude,heading,speed and vertical speed features contained in the historical track data,it was found that there is a strong correlation between changes in the above-mentioned feature values and changes in track trends,which can be used as input features for the model.Then,a DTW-based aggregated hierarchical clustering model including longitude,lati-tude and altitude three-dimensional location features is constructed to address the problems of existing trajectory clustering and single-method prediction models considering only two-dimensional location features.And on this basis,four trajectory prediction models based on Stacked LSTM,Bidirectional LSTM,CNN LSTM and Conv LSTM were built.Through the trajectory clustering of historical tracks,the obtained classification results can be used to cal-culate a prior knowledge related to the current track category in the prediction of the track,and add it to the prediction model as a feature to improve the prediction accuracy.Finally,the experimental environment is constructed to complete the simulation validation of the clustering and prediction models.Normalize and standardize the track sample set,and construct the cluster model sample set and the prediction model sample set according to the input conditions of the cluster model and the prediction model.According to the output results of the clustering model and the prediction model,two clustering model evaluation indicators,DBI and DVI,and three prediction model evaluation indicators,horizontal,vertical,and time,were proposed.The clustered sample sets were input into the two clustering models,and the recorded experimental results were analyzed to find that the DBI values in the 3D clustering re-sults were smaller and the DVI values were larger compared to the 2D track clustering,proving that the 3D track clustering has more accurate track classification than the 2D track clustering.The clustering results and prediction sample sets were input to the prediction model for multi-ple comparison experiments.By comparing the prediction results on the test set,it was found that the Bidirectional LSTM-based and Conv LSTM-based trajectory prediction models after inputting the 3D clustering results had smaller error values corresponding to each index,which proved that the accuracy of the trajectory prediction models based on trajectory clustering and LSTM proposed in this paper was higher than that of the comparison models.