Research On AIS Trajectory Prediction Method Based On Machine Learning

In the context of the globalization of the world economy,shipping has been one of the most important forms of transportation in international trade,and the global shipping industry bears about 90% of world trade volume.With the rapid development of the world economy,the shipping industry has shown significant changes.The number and types of ships are continually increasing.New routes are continually emerging,and many hot waters has emerged as the times require.Although the increase in the number of ships and shipping routes has made shipping trade more and more prosperous,it will also crowd the channels in hot water and increase the load.Correspondingly,due to the ship’s problems and human factors,accidents will increase,which seriously threatens the lives and property safety of ship personnel.Therefore,the prediction of future ship trajectories through historical ship trajectory data has become the key to ensure the safe navigation of ships in the waters.With the promotion and popularization of the Automatic Identification System(AIS)system,the availability of ship AIS trajectory data is improved,and the AIS data that can be collected is more abundant,providing a conditional basis for ship trajectory prediction.This article analyzes and summarizes the relevant characteristics and structure of ship AIS trajectory data.Using machine learning related methods,a ship trajectory prediction model is set up to predict the ship’s future trajectory.The main research work of this paper includes:(1)Research on data completion and exception handling methods.Based on AIS raw data,data completion and abnormal data processing have to be carried out.(2)Research on clustering re-regression method for ship trajectory prediction.Combined with the classification idea,the k-medoids method is used to cluster the trajectory samples,and regression prediction is made in each class,which effectively reduces the differences between the trajectory samples.The experimental results demonstrate that the clustering re-regression method can improve the prediction accuracy.(3)Research on improving Hausdorff trajectory similarity measurement method.Founded on the principle of the traditional Hausdorff distance measurement method,this paper proposes an improved Hausdorff distance measurement method.The improved method is changed from the point-to-point calculation of the customary method to the point-to-line distance calculation.The experimental verification proves that the improved measurement method can effectively improve the clustering quality.(4)Research on model integration method.Combined with stacking idea and weighted integration idea,this paper proposes a model integration method based on XGBoost,Light GBM and LSTM,which increase the difference between models and enhances the generalization ability.Experimental results demonstrate that the model integration algorithm can effectively improve the prediction accuracy.