Research On The Method Of Track Sequence Prediction Based On Ship Trajectory

With the vigorous development of maritime industry,ocean transportation has occupied an important position in the world’s logistics system.The substantial increasement of trade volume has resulted ship density growth in the hot waters and channel.It is imperative to supervise,manage ships and avoid maritime traffic accidents.Trajectory prediction is the basis of other research directions such as ship anomaly detection and collision avoidance.Accurately and efficiently predicting the future trajectory of ships is of great significance for reducing the incidence of maritime accidents and improving the management level of regulatory agencies.Aiming at the issue of trajectory prediction,this paper builds the data processing model to process a large amount of raw AIS data and convert the raw AIS data to trajectory sequence,which involves the steps of decoding,data cleaning,trajectory extraction and interpolation processing of missing value.After that,BP neural network,extreme learning machine and long short-term memory model are used to construct the track prediction model to complete the task of track prediction.The experimental results show that the prediction model based on BP neural network,extreme learning machine and long short-term memory model can generally achieve high performance in the prediction of longitude,latitude,speed and heading.The prediction accuracy and stability of BP network are worse than extreme learning machine and long short-term memory model due to its shortcomings of gradient vanishment and easy to fall into local optimal solutions.The long short-term memory model can effectively solve the problem of long-term dependence in the time dimension which leads to the best performance.The extreme learning machine is slightly better than the BP network in performance.The advantage of the extreme learning machine is that it can greatly reduce the training time,and has no problems of gradient disappearance and explosion.The long short-term memory model can capture the dependency relationship between successive track points in the time dimension.Comparing with the BP network and the extreme learning machine,the performance has significantly improved.However,the long short-term memory model ignores the relationship among attribute variables which is called spatial dependence.In order to further improve the prediction performance of the trajectory prediction model,this paper tries to use the graph neural network with the powerful ability in representing spatial dependence to construct the prediction model.Graph learning module is designed to solve the problem of uncertain external graph structure and adaptive change of graph structure.A graph neural network framework that can be used to process the track sequence is constructed.The experimental results show that the constructed graph neural network model can be used for trajectory prediction,and due to comprehensive consideration of the time dependence and spatial dependence of multivariate time series,the prediction performance of the model is better than the long short-term memory model.In the prediction task for longitude,latitude,speed and heading,the total MAE of the four features have declined by 5.65%,and the prediction stability was also significantly better than the long shortterm memory model.Finally,a comparative experiment on the structure of the model proved the rationality of the structure of the graph neural network framework constructed in this paper.