Study On Vessel Trajectory Prediction Model Based On AIS Data

Marine transportation has become one of the most important methods of international commodity exchange.The total freight volume accounts for about 80% of the total international transportation volume and is still increasing rapidly.In this case,the maritime traffic routes are extremely complicated,and frequent traffic accidents will bring huge economic losses and hidden dangers to personnel safety.For oceangoing vessels,the significance of trajectory prediction lies in the ability to effectively control fuel consumption to achieve greater profits and avoid entering dangerous waters and affecting the timeliness of freight;for offshore vessels,the significance of trajectory prediction is to provide the owner with collision avoidance and routes Technical support for planning decisions.In recent years,the analysis and prediction of ship trajectory data has attracted more and more attention from scholars.At present,most of the existing methods are based on traditional machine learning methods or ship kinematics equations.There are also a few methods based on deep learning models but the prediction time is short.These models have been difficult to adapt to the increasingly complex maritime traffic conditions.Nowadays,almost all vessels are equipped with automatic identification system(AIS).AIS equipment can continuously broadcast the dynamic and static information of the ship to the outside world.Compared with the traditional Vessel Traffic Service(VTS)data and radar data,AIS data is easier to obtain.Based on a large amount of historical AIS data and without destroying its data structure,this paper conducts trajectory extraction.For the extracted data set,a ship trajectory prediction model with relatively long-term prediction and multi-dimensional prediction capabilities is established based on the deep learning model.The research content of this article is as follows:(1)Time feature is an important feature of trajectory data.In order to avoid structural damage to AIS data caused by various interpolation methods,this paper used a new trajectory division method,which includes two steps: the first step is to calculate the time increments between two adjacent points,the second step is to analyze the distribution of time increments,and determine the time increment trajectory division threshold to divide the trajectory.The trajectory obtained by division is still relatively complicated and cannot be used directly as a data set,and sub-trajectory extraction operations are required.In order to model the trajectory more reasonably,the final data set should contain a certain amount of curved trajectories and these trajectories should not be too complicated.This requires a reasonable extraction length of the subtrajectory.Therefore,this paper conducted an experiment to determine the length of the sub-trajectory extraction.(2)For avoiding the influence of the same trajectory shape in different positions,this paper used the longitude increment and latitude increment of each point to replace the original longitude and latitude information.The paper used five characteristics of trajectory including time increment,longitude increment,latitude increment,speed over ground and course over ground to make the purpose of multi-dimensional prediction.(3)Deeply studied the principle of Long Short-Term Memory(LSTM)and its advantages over Recurrent Neural Network(RNN).The influence of the parameters of the particle swarm optimization(PSO)on its performance and the improvement strategy are analyzed in detail,and the PSO-LSTM ship trajectory prediction model based on the improved particle swarm optimization is established.Experiments show that the predictive ability of the PSO-LSTM model is better than that of the pure LSTM model.(4)In order to further improve the relative long-term prediction ability of the model,a vessel trajectory prediction model based on variational LSTM(VLSTM)is established.Experiments show that the prediction model based on VLSTM has shown better capabilities in relatively long-term prediction and curve prediction.