Research On Port Vessel Trajectory Prediction Method Based On AIS Data

The Automatic Identification System(AIS)is a system that uses time division multiple access technology in the VHF maritime mobile frequency band,which can automatically broadcast and receive dynamic and static information related to ship voyages.The rapid development of the AIS makes it possible to obtain massive amounts of vessel trajectory data,and provides basic conditions for vessel trajectory data mining,especially the research of vessel trajectory prediction.Based on the historical AIS trajectory data and the deep learning method,the paper fully excavates the hidden ship motion pattern behind the data,and realizes the prediction of the navigation trajectory of the port vessel in the next five moments.The prediction features include time difference,latitude and longitude,speed over ground and course over ground,which have important application value for vessels to enter and exit ports safely and efficiently in complex navigation environments,and for maritime departments to achieve efficient traffic management.The main contents of the paper are as follows:(1)AIS trajectory data preprocessing.Extract the vessel trajectory data near the Port of New York and New Jersey from the AIS data.Then,based on the combination of the AIS technical characteristics,the basic principle of kinematics and threshold method,the abnormal data of the time difference,latitude,longitude,speed over ground and course over ground are detected,eliminated and corrected.Finally,the vessel trajectory is segmented and sub-trajectory extraction is performed based on the time difference threshold,then,a trajectory prediction data set is constructed.In addition,the stationarity of the sub-trajectory is analyzed to provide a basis for the selection of the trajectory prediction model.(2)Construct two port vessel trajectory prediction models respectively based on typical Long Short-Term Memory(LSTM)and Temporal Convolutional Networks(TCN).Both above methods realize the multi-feature trajectory prediction at multiple times in the future through multi-step iterations.TCN,as a new model for processing time series,carries out feature extraction and long-time dependent information mining of trajectory through its powerful receptive field of convolution layer,and realizes prediction task in parallel and efficiently.The advantages of TCN model are confirmed by comparing its experimental results with that of LSTM model.(3)Construct a port vessel trajectory prediction model based on tensor train decomposition and LSTM(TLSTM).The TLSTM uses its high-order dynamics explicit modeling capabilities and long-term time dependence capture ability to achieve efficient and accurate sequence-to-sequence direct multi-step trajectory prediction.This paper discusses the effect of the number of LSTM layers and the number of hidden layer nodes on the TLSTM,and compares it with the Seq2 Seq model,which uses LSTM as the encoder and decoder,and the iterative prediction LSTM and TCN models.This paper also constructs an improved TLSTM vessel trajectory prediction model based on Distortion Loss including sh Ape and Tim E(DILATE).The DILATE loss function for model training can improve the shape distortion and time distortion of the predicted trajectory sequence to some extent.Moreover,the effectiveness of the DILATE＿TLSTM method is verified by comparing with the TLSTM.The experimental results show that,compared with the LSTM model with good trajectory prediction effect before,TCN performs lower mean square error and higher accuracy for each feature prediction of port vessel trajectory at the next five moments.However,the TCN model also has certain limitations,its prediction error will increase with the prediction time,which is not suitable for long-term trajectory prediction.The prediction effect of the TLSTM model based on the direct multi-step prediction is better than that of Seq2 Seq and TCN,and there is no accumulation of iteration errors,which is more suitable for long-term trajectory prediction.While the mean square error of the DIALTE＿TLSTM model is not lower than that of the TLSTM,it reduces the time distortion index and improves the trajectory fitting effect.Overall,DILATE＿TSLTM model has the best prediction performance,and it has good applicability for the prediction of port vessels trajectory at multiple consecutive times,which provides reliable support for port safety and efficient management.