CN113240199B

CN113240199B - Port ship track prediction method based on DILATE _ TLSTM

Info

Publication number: CN113240199B
Application number: CN202110630985.6A
Authority: CN
Inventors: 吴尽昭; 苏伟; 张久文; 蔡川; 雷晖; 刘映杰
Original assignee: Lanzhou University; Guangxi University for Nationalities
Current assignee: Lanzhou University; Guangxi University for Nationalities
Priority date: 2021-06-07
Filing date: 2021-06-07
Publication date: 2022-10-28
Anticipated expiration: 2041-06-07
Also published as: CN113240199A

Abstract

The invention relates to a port ship track prediction method based on DILATE _ TLSTM, which specifically comprises the steps of S1, collecting port AIS data, preprocessing the collected data and obtaining a track prediction data set; s2, constructing a port ship track prediction model based on DILATE _ TLSTM; s3, processing the sequence of the input model by adopting a sliding window method, and training the model to obtain a prediction result; and S4, carrying out measurement and evaluation on the prediction result by using the average absolute error, the mean square error, the goodness of fit, the shape distortion DTW and the time distortion index TDI to obtain the prediction effect of the port ship track prediction model. The port ship track prediction model based on DILATE _ TLSTM constructed by the invention can realize direct multi-step prediction, and improves the TLSTM model based on the shape and time distortion loss function, thereby improving the fitting capability of the track sequence prediction model.

Description

Port ship track prediction method based on DILATE _ TLSTM

Technical Field

The invention relates to the technical field of intelligent prediction of ship tracks, in particular to a port ship track prediction method based on DILATE _ TLSTM.

Background

Shipping is the main transportation form of international trade, and under the condition of economic globalization, the connections among countries are becoming more and more compact, and marine transportation becomes an important way for trading among countries, and the amount of marine trading accounts for about 90% of the total world trade amount and keeps higher than the increase rate of global economy. Along with the rapid development of shipping industry and the acceleration of port production and construction, the more ships are made, the more the varieties are increased, the number of the ships is increased continuously, the prosperity of maritime trade is brought, the potential safety hazard is brought to water traffic, the collision probability of the ships in the shipping process is higher and higher, and the safety of ship transportation is seriously influenced. According to statistics, coastal and offshore water areas are areas with frequent occurrence of marine accidents, 90% of marine accidents occur in the offshore water areas, particularly near ports, the port traffic safety management conditions become increasingly complex, for ship owners, direct support is provided for ship collision avoidance early warning and route planning decision by acquiring the navigation routes of ships in advance, and safety accidents can be effectively prevented. In the aspect of port scheduling, along with the increasing number of ships, the scheduling management of the ships at the ports faces huge examination, the difficulty of the ship scheduling is gradually increased due to the complex condition of large-scale ports, and dynamic information of the ships needs to be acquired urgently to formulate an effective port scheduling plan; in the aspect of port piloting, because of the complexity of the water environment and the density of port ships, the sailing dynamic state of other ships is known in advance, and the piloting plan is adjusted in time, so that the method has great significance for piloting the ships to enter and exit the port on time and improving the piloting work efficiency. Therefore, the navigation dynamic state of the ship is mastered and predicted in real time, and necessary support can be provided for port traffic management, port ship scheduling coordination and avoidance assessment schemes.

Disclosure of Invention

The invention aims to construct a port ship track prediction model based on DILATE _ TLSTM, realize direct multi-step prediction and improve the fitting effect of track prediction.

In order to achieve the purpose, the invention provides the following scheme:

a port ship track prediction method based on DILATE _ TLSTM comprises the following steps:

s1, collecting port AIS data and preprocessing the collected data to obtain a track prediction data set;

s2, constructing a port ship track prediction model based on DILATE _ TLSTM based on the acquired available track prediction data set;

s3, processing a sequence input into a DILATE _ TLSTM model by adopting a sliding window method, and then training the DILATE _ TLSTM model to obtain a prediction result;

and S4, carrying out measurement and evaluation on the prediction result by using the average absolute error, the mean square error, the goodness of fit, the shape distortion DTW and the time distortion index TDI to obtain the prediction effect of the port ship track prediction model.

Preferably, the data preprocessing comprises track extraction, abnormal value processing and track segmentation, and the preprocessed data is subjected to normalization processing and ship type characteristic one-hot coding to obtain a model training data set and data sequence stationarity analysis to obtain a stationarity analysis result of the training data set.

Preferably, the abnormal value processing is based on AIS technical characteristics, a mode of combining a basic principle of kinematics with a threshold value method, and the abnormal data of time interval, longitude and latitude, speed to ground and heading are detected, removed and corrected.

Preferably, the specific steps of constructing the port ship trajectory prediction model based on the DILATE _ TLSTM include:

s1.1, firstly, constructing a TLSTM-based port ship track prediction model: determining input characteristics as time interval, longitude and latitude, navigational speed, course and ship type, and output characteristics as time interval, longitude and latitude, navigational speed and course, and establishing a TLSTM-based port ship track prediction model according to a stationarity analysis result of the training data set and a multi-step time dependency relationship between the input characteristics and the output characteristics;

s1.2, taking ship track characteristics X (t-L + 1), \ 8230, X (t-1) and X (t) at continuous L moments as input data of a model, processing the input data, and directly predicting and outputting track characteristics Y (t + 1), Y (t + 2) \8230andY (t + N) at subsequent N moments;

s1.3, training the TLSTM model by taking DILATE as a loss function to obtain a port ship track prediction model based on DILATE _ TLSTM.

Preferably, an encoder and a decoder in the DILATE _ TLSTM model both adopt a two-layer LSTM structure, and the encoder is used for encoding a ship historical track and extracting a motion mode of a ship in a dynamic environment; the decoder obtains the ship motion mode characteristics by using the encoder and predicts the future motion trail of the ship.

Preferably, the output characteristics of the DILATE _ TLSTM model are five characteristics of time interval, latitude, longitude, speed and heading.

Preferably, the activation function of the DILATE _ TLSTM model adopts a Leaky ReLU function, and the optimizer adopts an Adam optimizer.

Preferably, the DILATE is used for loss evaluation of two time series, and mutation information and distortion information on shapes are captured, so that the sequence fitting effect is improved.

Preferably, the DILATE _ TLSTM model is constructed based on a Tensor Train decomposition and an Encoder-Decoder framework for reducing the number of model parameters using Tensor training.

Preferably, in S3, the evaluation index further includes shape distortion DTW and time distortion TDI for further evaluating the trajectory fitting effect.

The invention has the beneficial effects that:

the port ship track prediction model based on DILATE _ TLSTM constructed by the invention can realize direct multi-step prediction, and solves the problem of error accumulation of a multi-step iterative prediction model due to the increase of prediction step length; and the TLSTM model is improved based on the shape and the time distortion loss function, so that the mutation information of the sequence and the time distortion phenomenon of sequence prediction can be effectively extracted, the fitting capability of the track sequence prediction model is further improved, and a more reliable basis is provided for efficient port management.

Drawings

In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings needed to be used in the embodiments will be briefly described below, and it is obvious that the drawings in the following description are only some embodiments of the present invention, and it is obvious for those skilled in the art to obtain other drawings without inventive exercise.

FIG. 1 is a flow chart of the method of the present invention;

FIG. 2 is a diagram of a ship trajectory prediction model structure of DILATE _ TLSTM port according to the present invention;

FIG. 3 is a plot of the loss of the Seq2Seq model training in the embodiment of the present invention;

FIG. 4 is a plot of TLSTM training loss in an embodiment of the present invention;

fig. 5 is a graph of mean square error distribution of different numbers of hidden layers and numbers of hidden layer nodes in the embodiment of the present invention;

FIG. 6 is a comparison graph of the actual fitting effect of the trajectory prediction of each model in the embodiment of the present invention.

Detailed Description

The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it is obvious that the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. All other embodiments, which can be obtained by a person skilled in the art without making any creative effort based on the embodiments in the present invention, belong to the protection scope of the present invention.

In order to make the aforementioned objects, features and advantages of the present invention comprehensible, embodiments accompanied with figures are described in further detail below.

A method for predicting the track of a port ship based on DILATE _ TLSTM (as shown in the attached figure 1) comprises the following steps:

s1, collecting port AIS data and preprocessing the collected data to obtain a track prediction available data set;

s2, constructing a port ship track prediction model based on DILATE _ TLSTM (as shown in the attached figure 2);

s3, processing the sequence of the input model by adopting a sliding window method, and then training the model to obtain a prediction result;

And in the further optimization scheme, the data preprocessing part comprises track extraction, abnormal value processing and track segmentation, and the processed data is subjected to normalization processing and ship type characteristic one-hot coding to obtain a model training data set and perform data sequence stability analysis.

According to the further optimization scheme, abnormal value processing is based on AIS technical characteristics, a basic kinematics principle and a threshold method, time interval, longitude and latitude, speed to ground and navigation speed and abnormal data of course to ground are detected, eliminated and corrected, and compared with single threshold method abnormal detection and linear interpolation method correction, the method can achieve maximum track restoration without changing an original AIS data structure.

Further optimizing the scheme, the specific steps of constructing the port ship trajectory prediction model based on DILATE _ TLSTM comprise:

s1.1, firstly, constructing a TLSTM-based port ship track prediction model: determining input characteristics as time interval, longitude and latitude, navigational speed, course and ship type, and output characteristics as time interval, longitude and latitude, navigational speed and course, and establishing a TLSTM-based port ship track prediction model according to the stability analysis result of the training data set and the multi-step time dependency relationship between the input characteristics and the output characteristics;

s1.2, ship track features X (t-L + 1), \8230, X (t-1) and X (t) at continuous L moments are used as input of a model, input data are processed into an input form suitable for a TLSTM model, (Sample, timestap, feature), the Sample represents a Sample, the Timestap represents a time step, namely the number of historical track points, the Feature represents attributes of the Sample, namely six input features, and then track features Y (t + 1), Y (t + 2) 8230, Y (t + N) at subsequent N moments are directly predicted and output;

According to a further optimization scheme, an encoder and a decoder in the DILATE _ TLSTM model both adopt a two-layer LSTM structure, and the encoder is used for encoding historical tracks of a ship and extracting a motion mode of the ship in a dynamic environment; the decoder obtains the ship motion mode characteristics by using the encoder and predicts the future motion trail of the ship.

In a further optimization scheme, the output characteristics of the DILATE _ TLSTM model are five characteristics of time interval, latitude, longitude, navigation speed and heading.

In a further optimization scheme, an activating function of the DILATE _ TLSTM model adopts a Leaky ReLU function, and an optimizer adopts an Adam optimizer.

And in a further optimization scheme, the DILATE is used for performing loss evaluation on the two time sequences, capturing shape mutation information and time distortion information, and improving the sequence fitting effect.

According to the further optimization scheme, the DILATE _ TLSTM model is constructed based on a transducer Train decomposition and an Encoder-Decoder framework, and the number of model parameters is reduced by using Tensor training.

In a further optimization scheme, in S3, the evaluation index further includes shape distortion DTW and time distortion TDI, which are used to further evaluate the trajectory fitting effect.

In order to verify the technical effect of the invention, in the embodiment, a port ship track prediction model based on Seq2Seq and TLSTM and a port ship track prediction model based on ditate _ TLSTM are compared and analyzed.

Seq2Seq Experimental Performance analysis

For a port ship track prediction model based on Seq2Seq, the division ratio of a training set, a verification set and a test set is 8:1:1, MSE is adopted in the experiment as a loss function, adam is adopted in an optimizer, and main hyper-parameters are determined preferentially from multiple experiments and are set as follows: the time step is 10, the batch size is set to 128, the learning rate is set to 1e-5, the epoch is set to 15, the number of LSTM hidden layer nodes is 64, and the training loss curve of the obtained model is shown in FIG. 3.

As can be seen from FIG. 3, the error of the verification set is recorded and lags behind the error of the training set to record one Epoch, so that at the beginning of the curve, the loss of the verification set is lower than that of the training set, both loss curves tend to converge within 4 epochs, and the model is reasonable after the training is finished.

TLSTM experimental performance and key parameter analysis

For a TLSTM-based port ship trajectory prediction model, the prediction task is still the same as the trajectory prediction of the next 5 moments, and the loss function, the optimizer and the Epoch number adopted by the division of the training set, the verification set and the test set and the experiment are all the same as those of the Seq2Seq model. Other main hyper-parameters of the model are preferentially determined from multiple experimental results and are set as follows: the time step L is 10, the batch size is set to be 128, the learning rate is set to be 1e-3, the TT rank is 2, the hysteresis order is set to be 2 according to the sequence stationarity analysis result, the number of LSTM layers and the number of LSTM hidden layer neurons are respectively set to be 2 and 64, and the training loss curve of the obtained model is shown in figure 4. As can be seen from FIG. 4, the loss curves of the training set and the verification set basically tend to converge within 2 epochs, and the model training effect is good as compared with the Seq2Seq model, wherein the convergence is faster.

Generally, the selection of the number of LSTM layers and the number of LSTM hidden layer neurons has a large influence on the model performance, and therefore the selection of the number of LSTM layers and the number of hidden layer nodes is discussed for predicting the mean square error. The LSTM layer number is single-layer or two-layer, and the LSTM hidden layer node number is 8, 16, 32, 64, 128 and 256, respectively, to obtain the mean square error distribution of the track prediction model with different LSTM layer numbers and different hidden layer node numbers, as shown in FIG. 5, as can be seen from FIG. 5, increasing the hidden layer node number can reduce the model error to a certain extent, but as the increase, the network is complicated, and the training effect is reduced; and under the condition that the number of nodes of the hidden layer is the same, the model error can also be reduced by increasing the number of LSTM layers, and the track prediction mean square error MSE reaches the lowest value of 0.0045097 when the number of LSTM layers is 2 and the number of nodes of the hidden layer is 64, so that the number of LSTM layers and the number of nodes of the hidden layer of the TLSTM model are finally determined to be 2 and 64 respectively.

Comparison and analysis of experimental results of TLSTM and Seq2Seq models

Respectively predicting the tracks on a test set by using the trained TLSTM model and the Seq2Seq model to obtain a sequence of five corresponding characteristics including delta Time, LAT, LON, SOG and COG, and using MSE, MAE and R ² Three evaluation indexes are used for evaluating and comparing the prediction effects of the two models on the five characteristic sequences, and the results are shown in table 1:

TABLE 1

As can be seen from the above table, TLSTM realizes high-precision direct multi-step prediction of latitude, longitude and heading to ground, and the coefficient R is determined by prediction ² Respectively reaches 0.9963441, 0.9974195 and 0.9280750, conforms to the actual requirements of ports to a high degree, and determines a coefficient R for direct multi-step prediction of the navigational speed prediction ² At 0.8416165, there is still room for improvement, and for the prediction of the time interval, the result shows that neither TLSTM nor Seq2Seq can solve the problem of ineffective prediction caused by uneven distribution of the time interval.

Actual error of TLSTM single feature prediction

When data is preprocessed, normalization processing is carried out on input features in order to eliminate the influence of dimension difference among the input features on model convergence, error evaluation of a prediction result is also based on the normalized data, and a specific error value of the model on single feature prediction cannot be visually seen, so that reverse normalization processing needs to be carried out on the predicted value, a mean square error and a maximum error are used for evaluating a reverse normalization predicted value of the single feature, and the result is shown in a table 2. As can be seen from table 2, the maximum errors of the TLSTM model for latitude and longitude predictions are only 0.015 degree and 0.011 degree, respectively, so that high-precision prediction is achieved, the maximum errors for time interval, navigational speed, and heading predictions are 0.8 section and 1.17 degree, respectively, the prediction errors are within an acceptable range of practical application, and when the prediction for time interval has a large difference between the time interval predicted at the next moment and the input time interval sequence, the error is large, and is 16 seconds. In general, the accuracy of each feature prediction can well meet the requirements of port ship management.

TABLE 2

5. Port ship track prediction experiment result analysis based on DILATE _ TLSTM

In a port ship track prediction experiment based on DILATE _ TLSTM, the same parameter setting as that of a TLSTM port ship track prediction model is adopted, and a shape distortion weight factor is alpha =0.5 to train the model. The trained model was evaluated using the MSE, DTW, TDI evaluation indices with the test set, and compared with the TLSTM model using the same evaluation indices, with the results shown in table 3:

TABLE 3

As can be seen from table 3 above, from the evaluation index MSE, although the prediction errors of the tstlm model based on the ditate loss function for five features are slightly higher than those of the TLSTM model based on the MSE loss function, the difference is extremely small, and in this evaluation level, the prediction performances of the two models can be considered to be equivalent. However, from the evaluation index TDI, the time distortion of the DILATE _ TLSTM model to the time interval, the latitude and the earth heading prediction is obviously reduced, which means that the fitting effect of the sequence can be improved to a certain degree under the condition that the MSE is basically the same.

According to the method, a Seq2Seq, TLSTM and DILATE _ TLSTM model is constructed from the perspective of sequence to sequence prediction, direct multi-step prediction of five trace point characteristics at five moments in the future is realized, and the Seq2Seq is used as a reference comparison model. Due to the fact that nonlinear sudden change and time distortion exist in ship trajectory prediction, a DILATE loss function aims at capturing shape distortion and time distortion of a trajectory sequence when a TLSTM model is trained so as to try to improve a prediction fitting effect. The experimental result analysis and the comparison with the experimental result of the Seq2Seq model show that the prediction precision of TLSTM and DILATE _ TLSTM models is higher in multi-step port ship track prediction, and the direct multi-step prediction model is more suitable for long-time track prediction; compared with the conventional MSE loss function, the DILATE loss function can capture information of sudden change and time distortion, reduces time distortion of a predicted track sequence, and improves the fitting effect of track prediction. The actual fit effect pairs of the three models to port vessel trajectory predictions are shown in fig. 6.

The port ship track prediction model based on DILATE _ TLSTM constructed by the invention can realize direct multi-step prediction, and solves the problem of error accumulation caused by the increase of prediction step length of a multi-step iterative prediction model; and the TLSTM model is improved based on the shape and the time distortion loss function, so that the mutation information of the sequence and the time distortion phenomenon of sequence prediction can be effectively extracted, the fitting capability of the track sequence prediction model is further improved, and a more reliable basis is provided for efficient port management.

The above-described embodiments are only intended to describe the preferred embodiments of the present invention, and not to limit the scope of the present invention, and various modifications and improvements made to the technical solution of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A port ship track prediction method based on DILATE _ TLSTM is characterized by comprising the following steps:

the specific steps of constructing the port ship track prediction model based on DILATE _ TLSTM comprise the following steps:

s2.1, firstly, constructing a TLSTM-based port ship track prediction model: determining input characteristics as time interval, longitude and latitude, navigational speed, course and ship type, and output characteristics as time interval, longitude and latitude, navigational speed and course, and establishing a TLSTM-based port ship track prediction model according to a stationarity analysis result of a training data set and a multi-step time dependency relationship between the input characteristics and the output characteristics;

s2.2, taking ship track characteristics X (t-L + 1), \8230, X (t-1) and X (t) at continuous L moments as input data of a model, processing the input data, and directly predicting and outputting track characteristics Y (t + 1), Y (t + 2), \8230, and Y (t + N) at subsequent N moments;

s2.3, training the TLSTM model by taking DILATE as a loss function to obtain a port ship track prediction model based on DILATE _ TLSTM;

s3, processing the sequence input into a DILATE _ TLSTM model by adopting a sliding window method, and then inputting into the DILATE _ TLSTM model to obtain a prediction result;

s4, carrying out measurement and evaluation on the prediction result by using the average absolute error, the mean square error, the goodness of fit, the shape distortion DTW and the time distortion index TDI to obtain the prediction effect of the port ship track prediction model;

the data preprocessing comprises track extraction, abnormal value processing and track segmentation, and the preprocessed data is subjected to normalization processing and ship type characteristic one-hot coding to obtain a model training data set and perform data sequence stability analysis to obtain a stability analysis result of the training data set;

the abnormal value processing is based on AIS technical characteristics, a basic kinematics principle and a threshold value method, and is used for detecting, eliminating and correcting time interval, longitude and latitude, speed to ground and heading abnormal data;

an encoder and a decoder in the DILATE _ TLSTM model both adopt a two-layer LSTM structure, and the encoder is used for encoding a historical track of a ship and extracting a motion mode of the ship in a dynamic environment; the decoder obtains the ship motion mode characteristics by using the encoder and predicts the future motion trail of the ship.

2. The DILATE _ TLSTM-based harbor vessel trajectory prediction method according to claim 1, characterized in that the activation function of the DILATE _ TLSTM model is a Leaky ReLU function, and the optimizer is an Adam optimizer.

3. The DILATE _ TLSTM-based harbor ship trajectory prediction method according to claim 1, characterized in that DILATE is used for loss evaluation of two time series, capturing information of sudden change in shape and distortion in time, and improving the sequence fitting effect.

4. The DILATE _ TLSTM-based port ship trajectory prediction method of claim 1, characterized in that the DILATE _ TLSTM model is constructed based on a temporal transform decomposition and an Encoder-Decoder framework for reducing the number of model parameters using Tensor training.