A self-correcting intelligent ship trajectory prediction method

Abstract

The application discloses a self-correcting ship trajectory intelligent prediction method, comprising the following steps: obtaining an original AIS data set, cleaning and repairing the original AIS data set; dividing the AIS data set into n sub-data sets, sampling at equal time intervals from the n sub-data sets respectively, and storing the sampled trajectory point data as trajectory data sets of different ships respectively; constructing and training an XGBoost model, obtaining predicted trajectory points of the ship according to the XGBoost model, training k groups of XGBoost models based on k prediction time intervals, and obtaining k groups of predicted trajectory points; setting a position prediction error threshold and a similarity threshold, calculating the similarity values of the predicted trajectory points and the trajectory points in the trajectory data sets of different ships, calculating the weighted prediction errors of x similar historical trajectory points, correcting the trajectory point data according to the weighted prediction errors, and taking the k groups of corrected predicted trajectory point data as the predicted ship trajectory. The influence of data loss or data anomaly is reduced, and the trajectory prediction accuracy is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of ship trajectory prediction, in particular to a self-correcting intelligent ship trajectory prediction method. BACKGROUND

[0002] China has abundant shipping resources, and accelerating the development of the maritime transportation industry is of great significance for building a modern comprehensive transportation system, adjusting and optimizing the industrial layout in coastal areas, promoting energy saving and emission reduction, and coordinated development of regional economy. However, maritime transportation is a recognized high-risk industry. With the development of domestic economy and world trade, the transportation is increasingly busy, the number of ships is increasing, the ships are becoming larger and more specialized, the ship navigation speed is continuously improving, and the number of dangerous goods transportation is increasing, resulting in increasing density of water transportation and deteriorating ship navigation environment, causing frequent water transportation accidents. Therefore, for the water transportation management department, a ship trajectory prediction technology is urgently needed to reasonably schedule ships in hot water areas and reduce the probability of maritime transportation accidents. The Automatic Identification System (AIS) is a new type of navigation aid developed to address the above problems. AIS can realize the dynamic reception and transmission of dynamic information such as latitude, longitude, speed, heading, etc. and static information such as Maritime Mobile Service Identify (MMSI), ship length, draft, etc. The rapid development of AIS has made it widely used in maritime navigation collision avoidance, maritime supervision, etc. and made ship trajectory prediction possible.

[0003] Currently, the methods for predicting ship trajectory mainly fall into two categories: kinematics modeling-based methods and machine learning-based methods. The former only considers the current trajectory of the ship without considering the historical trajectory, ignoring the prior information of the current water area. The latter has high requirements for the scale and quality of ship trajectory data. SUMMARY

[0004] The present application provides a self-correcting intelligent ship trajectory prediction method to overcome the above technical problems.

[0005] A self-correcting intelligent ship trajectory prediction method, comprising,

[0006] S1, obtaining an original AIS data set, cleaning the original AIS data set, the cleaning including deleting abnormal AIS data, repairing and supplementing missing values in the cleaned AIS data set, the repairing and supplementing including repairing and supplementing according to the adjacent AIS data of the missing values by interpolation method;

[0007] S2, dividing the AIS dataset into n subsets by the maritime mobile service identity code in the AIS dataset, wherein n is the number of maritime mobile service identity codes in the AIS dataset, setting a time interval, sampling the n subsets at equal time intervals, respectively, and storing the sampled trajectory point data as trajectory datasets of different ships, wherein the trajectory point data at least includes longitude, latitude, speed, longitude change, and latitude change, and the longitude change and latitude change of the trajectory point are calculated according to the trajectory point data at adjacent time points;

[0008] S3, constructing a first XGBoost model based on longitude change and a second XGBoost model based on latitude change, inputting the longitude change, speed, and heading of the trajectory point in the trajectory dataset into the first XGBoost model and training, inputting the latitude change, speed, and heading of the trajectory point in the trajectory dataset into the second XGBoost model and training, inputting a single trajectory point into the two models respectively, obtaining the output longitude change and latitude change, and obtaining the longitude and latitude of the predicted trajectory point according to the longitude, latitude, output longitude change, and latitude change of the trajectory point;

[0009] S4, setting k consecutive prediction time intervals, inputting the longitude change, latitude change, speed, and heading of the i-th predicted trajectory point into the first XGBoost model and the second XGBoost model, respectively, obtaining the output longitude change and latitude change, and calculating the longitude and latitude of the i+1-th predicted trajectory point according to the longitude of the i-th predicted trajectory point, the latitude of the i-th predicted trajectory point, the longitude change, and the latitude change, wherein i≥1;

[0010] S5, setting a position prediction error threshold and a similarity threshold, judging whether the position of the i-th predicted trajectory point meets the position prediction error threshold, executing S6 when the position prediction error threshold is met, and otherwise calculating the similarity value of the i-th predicted trajectory point with the trajectory points in the trajectory datasets of different ships, obtaining x historical trajectory points that meet the similarity threshold, calculating the weighted prediction error of the x historical trajectory points, modifying the i-th trajectory point data according to the weighted prediction error, and obtaining the modified i-th predicted trajectory point data;

[0011] S6, i=i+1, repeating steps S4 and S5 until k sets of predicted trajectory point data are obtained, and taking the k sets of predicted trajectory point data as the predicted ship trajectory.

[0012] Preferably, S3 can further include normalizing the data input into the first XGBoost model and the second XGBoost model according to formula (1),

[0013]

[0014] wherein I represents the data inputted into the model, I' represents the normalized data, max[I] and min[I] represent the maximum and minimum values of I, respectively.

[0015] Preferably, the calculating the similarity value of the i-th predicted trajectory point with the trajectory points in the trajectory data set of the different ship comprises calculating the heading similarity of the predicted trajectory point with the trajectory points according to formula (2), calculating the ship speed similarity of the predicted trajectory point with the trajectory points according to formula (3), and calculating the position similarity of the predicted trajectory point with the trajectory points according to formula (4),

[0016] d c = |COG1-COG2| (2)

[0017] d s = |SOG1-SOG2| (3)

[0018]

[0019] wherein d c represents the heading similarity, d s represents the ship speed similarity, and d h represents the position similarity, COG1 represents the heading of the predicted trajectory point, COG2 represents the heading of the trajectory point, SOG1 represents the ship speed of the predicted trajectory point, SOG2 represents the ship speed of the trajectory point, dlat represents the latitude of the predicted trajectory point, dlon represents the longitude of the predicted trajectory point, lat1 represents the latitude of the trajectory point, and lon1 represents the longitude of the trajectory point,

[0020] d c is normalized according to formula (5), d s is normalized according to formula (6), and d h is normalized according to formula (7),

[0021]

[0022]

[0023]

[0024] wherein mind c represents the minimum value of d c , maxd c represents the maximum value of d c , mind s represents the minimum value of d s , maxd s represents the maximum value of d s , mind h represents the minimum value of dh the minimum value of d h d h the maximum value of d c , d s , and d h represent normalized heading, speed, and speed similarity values, respectively,

[0025] The similarity value of the predicted trajectory point and the trajectory point is calculated according to formula (8),

[0026] D = w c · d c + w s · d s + w h · d h (8)

[0027] where w c , w s , and w h are weights, and D is the similarity value of the predicted trajectory point and the trajectory point.

[0028] Preferably, the weighted prediction error of the x historical trajectory points is calculated by calculating the error weight of the jth historical trajectory point according to formula (9),

[0029]

[0030] where D j represents the similarity value of the jth historical trajectory point, and X is the number of historical trajectory points,

[0031] The ith trajectory point data is corrected according to formula (10),

[0032]

[0033] where w j is the error weight, is the prediction error of the jth similar point at the tth time, and y t represents the preliminary prediction result at the tth time, and y t represents the corrected result.

[0034] Preferably, the position prediction error threshold is set to μ + 2σ, where μ and σ represent the mean and deviation of the prediction error of the XGBoost model on the training set.

[0035] The application provides a self-correcting ship trajectory intelligent prediction method, which is based not only on current ship trajectory information, but also on historical AIS data, and through preprocessing of the AIS data, the influence of data loss or data anomaly within a certain range is reduced; meanwhile, the machine learning algorithm is used to correct the error of the ship trajectory prediction result, improve the trajectory prediction accuracy, and achieve a compromise between the trajectory prediction accuracy and the data accuracy requirement, so as to be more suitable for actual application scenarios. BRIEF DESCRIPTION OF DRAWINGS

[0036] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or the prior art description. Obviously, the drawings in the following description are some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0037] Figure 1 is a method flowchart of the present application;

[0038] Figure 2 is an AIS data cleaning principle of the present application;

[0039] Figure 3 is a ship trajectory prediction model based on XGBoost of the present application;

[0040] Figure 4 is a ship trajectory correction model based on XGBoost of the present application. DETAILED DESCRIPTION

[0041] In order to make the purpose, technical solutions and advantages of the embodiments of the present application more clear, the technical solutions in the embodiments of the present application will be described clearly and completely in combination with the drawings in the embodiments of the present application. Obviously, the described embodiments are some embodiments of the present application, not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the protection scope of the present application.

[0042] Figure 1 is a method flowchart of the present application, as shown in Figure 1 , the method of the present embodiment can include:

[0043] S1, obtaining an original AIS data set, the AIS data including dynamic data and static data. The AIS dynamic data includes a turning rate, a type, a positioning time, a longitude, a latitude, a heading, a ship heading, a speed, a navigation state, a positioning accuracy, a receiving time. The AIS static data includes a ship MMSI number, an IMO number, a call sign, a ship name, a ship type, a ship length, a ship width, a positioning type, an expected arrival time, a draft, a destination port, a device type, a receiving time, a positioning antenna to ship bow distance, a positioning antenna to ship stern distance, a positioning antenna to port side distance, a positioning antenna to starboard distance.

[0044] The dynamic data is broadcasted once every 2-10 seconds during navigation, and the static data is broadcasted once every 6 minutes by default, so the AIS data can restore the motion of the ship in the water area with high accuracy. However, the process of obtaining AIS data needs to go through sensor data acquisition, message generation, packaging, modulation, transmission, demodulation, decoding and other processes, which may have errors in time, space and other aspects, which will affect the accuracy and reliability of ship trajectory prediction. Therefore, the AIS data needs to be cleaned before data analysis and mining.

[0045] The original AIS data set is cleaned, and the cleaning includes deleting abnormal AIS data, and the AIS data cleaning principle is as shown in Figure 2 Firstly, based on the dynamic data and the static data, the abnormal processing and cleaning of the AIS data are realized. In the fields of the AIS static data, the MMSI is 9 digits according to international regulations; the ship length and the ship width are manually input, and the unit is meter; the draft information is manually modified with the ship voyage, and the unit is decimeter; the receiving time is the time when the AIS receiving equipment receives the AIS information. In the fields of the AIS dynamic data, the speed (speed over ground) is in units of 1 / 10 knots; the longitude and latitude are in units of 1 / 10000 minutes; the heading (true heading) has a value range of 0-359; the time stamp is the UTC second of the report generation. The data that does not conform to the standard in each attribute field is deleted. Further, based on the semantic information of MMSI in the static data and longitude, latitude, ship speed in the dynamic data, AIS information that is not in the target sea area and has abnormal speed is deleted.

[0046] The missing values in the cleaned AIS data set are repaired and supplemented, and the repairing and supplementing includes repairing and supplementing the missing values according to the adjacent AIS data by interpolation method, that is, extracting the MMSI, longitude and latitude, heading and speed and other features in the AIS data that are useful for ship trajectory prediction, and performing interpolation processing on the missing values in the adjacent information to realize data repair and supplement;

[0047] S2, divide the data set into n sub-data sets by the maritime mobile service identity code in the AIS data set, where n is the number of maritime mobile service identity codes in the AIS data set, set a time interval, and sample the n sub-data sets at equal time intervals, respectively, and store the sampled trajectory point data as trajectory data sets of different ships, the trajectory point data at least including longitude, latitude, speed, longitude change, and latitude change, and the longitude change and latitude change of the trajectory point are calculated according to the trajectory point data at adjacent time points, for a ship, the trajectory feature at t time can be expressed as Y(t)={lat,lon,v,c}, where lat,lon,v,c are the latitude, longitude, speed and heading of the ship at t time. For each independent trajectory point, the corresponding latitude and longitude change can be calculated by formulas (1) and (2):

[0048] Δlat=lat j -lat i (1)

[0049] Δlon=lon j -lon i (2)

[0050] Wherein, lat j represents the latitude of the jth trajectory point, lat i represents the latitude of the ith trajectory point, lon j represents the longitude of the jth trajectory point, and lon i represents the longitude of the ith trajectory point, and the ith trajectory point and the jth trajectory point are adjacent in time.

[0051] Ship trajectory prediction is essentially a regression problem using a machine learning model, taking the historical trajectory features of the ship as the model input and the ship trajectory feature data at a future time as the output, establishing a mapping relationship between the historical trajectory of the ship and the future trajectory feature data of the ship, and calculating and predicting the future sailing trend of the ship.

[0052] S3, because the displacement in the longitude direction and the latitude direction when the ship moves is not related, in order to better learn the nonlinear relationship in the trajectory data set, a first XGBoost model based on longitude change (longitude displacement) and a second XGBoost model based on latitude change (latitude displacement) are constructed, and the XGBoost-based ship trajectory prediction model is as shown in Figure 3 .

[0053] The XGBoost model is composed of N decision trees, and the optimization objective function is as shown in formula (3):

[0054]

[0055] y' = f(x) = ∑Nt=1ft(x) (1) i is the final prediction of the model, N represents the number of generated decision trees, ft(x) represents the prediction of the tth tree, and x represents the input sample. k is the final prediction of the model, N represents the number of generated decision trees, ft(x) represents the prediction of the tth tree, and x represents the input sample. k is the final prediction of the model, N represents the number of generated decision trees, ft(x) represents the prediction of the tth tree, and x represents the input sample. i is the final prediction of the model, N represents the number of generated decision trees, ft(x) represents the prediction of the tth tree, and x represents the input sample.

[0056]

[0057] where the first term is a differentiable loss function loss, is the prediction value calculated by the first t-1 trees. The second is the regularization term Ω(f t ), which can control the complexity of the model to avoid overfitting. Ω(f t ) can be written as shown in equation (5):

[0058]

[0059] where γ is the regularization coefficient, T is the number of leaf nodes, and XGBoost model is composed of N decision trees. The optimization objective function is shown in equation (6):

[0060]

[0061] y' = f(x) = ∑Nt=1ft(x) (1) i is the final prediction of the model, N represents the number of generated decision trees, ft(x) represents the prediction of the tth tree, and x represents the input sample. k is the final prediction of the model, N represents the number of generated decision trees, ft(x) represents the prediction of the tth tree, and x represents the input sample. k is the final prediction of the model, N represents the number of generated decision trees, ft(x) represents the prediction of the tth tree, and x represents the input sample. i is the final prediction of the model, N represents the number of generated decision trees, ft(x) represents the prediction of the tth tree, and x represents the input sample.

[0062]

[0063] where the first term is a differentiable loss function loss, is the prediction value calculated by the first t-1 trees. The second is the regularization term Ω(f t ), which can control the complexity of the model to avoid overfitting. Ω(f t ) can be written as shown in equation (8):

[0064]

[0065] where γ is the regularization coefficient, T is the number of leaf nodes, and w is the score on each leaf.

[0066] Taylor expansion of (7) gives a new representation of the objective function as shown in equation (9):

[0067]

[0068] where g k and h k are the first and second order gradient statistics of the loss function loss as shown in equation (10):

[0069]

[0070] The longitude change, speed, and heading of the trajectory points in the trajectory data set are input into the first XGBoost model and trained, and the latitude change, speed, and heading of the trajectory points in the trajectory data set are input into the second XGBoost model and trained. A single trajectory point is input into the two models respectively, and the output longitude change and latitude change are obtained. The longitude and latitude of the predicted trajectory point are obtained according to the longitude and latitude of the trajectory point and the output longitude change and latitude change.

[0071] However, only the coordinates of one time point are not enough to form a complete ship trajectory. It is the score on each leaf, therefore, a set of XGBoost models need to be trained to obtain the complete ship trajectory prediction.

[0072] S4, set k consecutive prediction time intervals, input the longitude change, latitude change, speed, and heading of the i-th predicted trajectory point into the first XGBoost model and the second XGBoost model respectively,

[0073] The inputs of the two models are respectively:

[0074] In the two matrices, each row represents the input at each time point, and a total of n time points are involved. Δlat i , Δlon i , v i , and c i represent the latitude change, longitude change, speed, and heading at time t n .

[0075] In order to avoid the influence of training data on the training of the model, and at the same time accelerate the convergence performance of the model, we perform normalization processing on the input data through equation (11), where I represents the input data of the network, I' represents the normalized data, max[I] and min[I] represent the maximum and minimum values of I.

[0076]

[0077] Where I represents the input data of the model, I' represents the normalized data, max[I] and min[I] represent the maximum and minimum values of I, respectively.

[0078] The information at each time is obtained by the recursive unit R of XGBoost, which means that in order to predict the position at multiple times, the output of the XGBoost model at the previous time is used as the input of the next time in the prediction process, that is, the prediction model established based on XGBoost is called all the time, and this prediction model is called a recursive unit. Assuming that t k The output of the time model is O x and O y , which can be represented by formula (12) and formula (13),

[0079] O x = R x (Δlat k , v k , c k ) (12)

[0080] O y = R y (Δlon k , v k , c k ) (13)

[0081] Where R x and R y represent the predicted recursive units in longitude and latitude, respectively, and Δlat k , Δlon k , v k , c k represent the input of the model, that is, the longitude and latitude change values, the speed, and the heading at the kth time.

[0082] The longitude change and the latitude change are obtained, and the longitude and the latitude of the i+1th predicted trajectory point are calculated according to the longitude of the ith predicted trajectory point, the latitude of the ith predicted trajectory point, the longitude change, and the latitude change, where i≥1.

[0083] After the future trajectory is preliminarily predicted, the next position of the target point is obtained, a threshold is set to determine whether the prediction error is abnormal and whether it can be corrected. The threshold is set to μ+2σ, where μ and σ represent the mean and variance of the prediction error of the corresponding model on the training set.

[0084] Once the prediction error exceeds this range, it is defined as abnormal, and the similarity between the target trajectory point and the historical trajectory points in the database is measured, mainly considering the heading, the ship speed, and the spatial position.

[0085] S5, set a position prediction error threshold and a similarity threshold, judge whether the position of the i-th predicted trajectory point meets the position prediction error threshold, and the ship trajectory correction model based on XGBoost is as shown in Figure 3

[0086] The heading similarity of the predicted trajectory point and the trajectory point is calculated according to formula (14), the ship speed similarity of the predicted trajectory point and the trajectory point is calculated according to formula (15), and the position similarity of the predicted trajectory point and the trajectory point is calculated according to formula (16),

[0087] d c = |COG1-COG2| (14)

[0088] d s = |SOG1-SOG2| (15)

[0089]

[0090] wherein d c represents the heading similarity, d s represents the ship speed similarity, and d h represents the position similarity, COG1 represents the heading of the predicted trajectory point, COG2 represents the heading of the trajectory point, SOG1 represents the ship speed of the predicted trajectory point, SOG2 represents the ship speed of the trajectory point, dlat represents the latitude of the predicted trajectory point, dlon represents the longitude of the predicted trajectory point, lat1 represents the latitude of the trajectory point, and lon1 represents the longitude of the trajectory point,

[0091] In order to measure the similarity of two points in terms of heading and ship speed, the absolute difference is calculated by formula (14) and (15). Obviously, the smaller the difference between the two, the more likely the two points are similar. In order to measure the similarity of two points in spatial position, Harstine distance is calculated by formula (16). The reason why the Euclidean distance is not directly applied is that the earth is an approximate sphere, and the Harstine distance takes into account the curvature.

[0092] d c is normalized according to formula (17), d s is normalized according to formula (18), and d h is normalized according to formula (19),

[0093]

[0094]

[0095]

[0096] wherein mind c ​min d c max d c min d c max d s min d s max d s min d s max d h min d h max d h max d h min d c min d s min d h min d

[0097] Since COG, SOG and the distance in spatial domain can represent the similarity of maneuvering behavior and the similarity of spatial position respectively, a hybrid weighted similarity is proposed to consider these two factors comprehensively, and the similarity value between the predicted trajectory point and the trajectory point is calculated according to formula (20),

[0098] D = w c · d c + w s · d s + w h · d h (20)

[0099] where w c , w s and w h are weights, and D is the similarity value between the predicted trajectory point and the trajectory point.

[0100] If the position prediction error threshold is met, S6 is executed, otherwise the similarity value between the i-th predicted trajectory point and the trajectory points in the trajectory data set of different ships is calculated, and x historical trajectory points meeting the similarity threshold are obtained. After this similarity measurement, the most similar K points are selected from the historical AIS data as the guidance for trajectory correction. In order to obtain the correction value of the initial predicted trajectory, the entire prediction process is performed on these K points, and since these K points are included in the historical AIS database, it means that the true value of the trajectory point is determined, and the corresponding prediction error can be easily obtained.

[0101] The weighted prediction error of the x historical trajectory points is calculated, and the i-th trajectory point data is corrected according to the weighted prediction error to obtain the corrected i-th predicted trajectory point data, specifically, the error weight of the j-th historical trajectory point is calculated by formula (21),

[0102]

[0103] wherein, D j represents the similarity value of the jth historical trajectory point, X is the number of historical trajectory points, the ith trajectory point data is corrected according to formula (22),

[0104]

[0105] wherein, w j is the error weight, is the prediction error of the jth similar point obtained after the jth similar point is predicted at the tth time according to the Xgboost model, y t represents the preliminary prediction result at the tth time, y′ t represents the corrected result.

[0106] S6, i = i + 1, steps S4 and S5 are repeatedly executed until k sets of predicted trajectory point data are obtained, and the k sets of predicted trajectory point data are taken as the predicted ship trajectory.

[0107] Overall beneficial effects: the present application provides a self-correcting intelligent ship trajectory prediction method, which is not only based on the current trajectory information of the ship, but also reduces the influence of data missing or data abnormality within a certain range by preprocessing AIS data based on historical AIS data; at the same time, the machine learning algorithm is used to correct the error of the ship trajectory prediction result, improve the trajectory prediction accuracy, and achieve a compromise between the trajectory prediction accuracy and the data accuracy requirement, so as to be more suitable for the actual application scene.

[0108] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present application, but not to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement for part or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present application.

Claims

1. A self-correcting intelligent prediction method of ship trajectory, characterized in that, The application relates to a ship trajectory prediction method and device. S1, acquiring an original AIS data set, cleaning the original AIS data set, wherein the cleaning comprises deleting abnormal AIS data, repairing and supplementing missing values in the cleaned AIS data set, wherein the repairing and supplementing comprises repairing and supplementing the missing values through an interpolation method according to adjacent AIS data of the missing values; S2, dividing the data set into n sub-data sets through marine mobile service identification codes in the AIS data set, wherein n is the number of marine mobile service identification codes in the AIS data set, setting a time interval, sampling the n sub-data sets at equal time intervals, respectively storing the sampled trajectory point data as trajectory data sets of different ships, and the trajectory point data at least comprising longitude, latitude, speed, heading, longitude change amount and latitude change amount, and calculating the longitude change amount and the latitude change amount of the trajectory point according to the trajectory point data at adjacent time points; S3, constructing a first XGBoost model based on the longitude change and a second XGBoost model based on the latitude change, inputting the longitude change amount, speed and heading of the trajectory point in the trajectory data set into the first XGBoost model and training the first XGBoost model, inputting the latitude change amount, speed and heading of the trajectory point in the trajectory data set into the second XGBoost model and training the second XGBoost model, inputting a single trajectory point into the two models, respectively obtaining the output longitude change amount and the output latitude change amount, and obtaining the longitude and the latitude of the predicted trajectory point according to the longitude, the latitude and the output longitude change amount and the output latitude change amount of the trajectory point; S4, setting k consecutive prediction time intervals, inputting the longitude change amount, the latitude change amount, the speed and the heading of the i-th predicted trajectory point into the first XGBoost model and the second XGBoost model, respectively, obtaining the output longitude change amount and the output latitude change amount, and calculating the longitude and the latitude of the i+1-th predicted trajectory point according to the longitude of the i-th predicted trajectory point, the latitude of the i-th predicted trajectory point, the longitude change amount and the latitude change amount, wherein i is greater than or equal to 1; S5, setting a position prediction error threshold and a similarity threshold, judging whether the position of the i-th predicted trajectory point meets the position prediction error threshold, executing S6 when the position prediction error threshold is met, and otherwise calculating the similarity value of the i-th predicted trajectory point and the trajectory points in the trajectory data sets of different ships, obtaining x historical trajectory points meeting the similarity threshold, calculating the weighted prediction error of the x historical trajectory points, correcting the i-th trajectory point data according to the weighted prediction error, and obtaining the corrected i-th predicted trajectory point data; S6, i = i + 1, repeatedly executing steps S4 and S5 until k groups of predicted trajectory point data are obtained, and taking the k groups of predicted trajectory point data as the predicted ship trajectory.

2. The self-correcting ship trajectory intelligent prediction method according to claim 1, characterized in that, The S3 comprises normalizing the data input into the first XGBoost model and the second XGBoost model according to formula (1). (1) wherein, denotes data inputted into the model, denotes normalized data, and respectively represent a maximum value and a minimum value of the data.

3. The self-correcting ship trajectory intelligent prediction method according to claim 1, characterized in that, The calculating the similarity value of the i-th predicted trajectory point and the trajectory points in the trajectory data set of different ships comprises calculating the heading similarity of the predicted trajectory point and the trajectory points according to formula (2), calculating the ship speed similarity of the predicted trajectory point and the trajectory points according to formula (3), and calculating the position similarity of the predicted trajectory point and the trajectory points according to formula (4), (2) (3) (4) wherein, represents a heading similarity, represents a speed similarity, represents a position similarity, represents a heading of a predicted trajectory point, represents a heading of a trajectory point, represents a speed of a predicted trajectory point, represents a speed of a trajectory point, represents a latitude of a predicted trajectory point, represents a longitude of a predicted trajectory point, represents a latitude of a trajectory point, represents a longitude of a trajectory point, According to formula (5), the normalized processing is performed on According to formula (6), the normalized processing is performed on According to formula (7), the normalized processing is performed on According to formula (8), the normalized processing is performed on (5) (6) (7) wherein denotes the minimum of denotes the maximum of denotes the minimum of denotes the maximum of denotes the minimum of denotes the maximum of , and denote the normalized heading, speed and speed similarity value, respectively, The similarity value of the predicted trajectory point and the trajectory points is calculated according to formula (8), (8) wherein, , and are weights, is a similarity value of the predicted trajectory point to the trajectory point.

4. The self-correcting ship trajectory intelligent prediction method according to claim 3, characterized in that, The calculating the weighted prediction error of the x historical trajectory points is respectively calculating the error weight of the j-th historical trajectory point by formula (9), (9) wherein, represents a similarity value of the jth historical trajectory point, X is the number of historical trajectory points, The i-th trajectory point data is corrected according to formula (10), (10) wherein, is the error weight, is the prediction error of the jth similar point at the tth time, represents the preliminary prediction result at the tth time, represents the corrected result.

5. The self-correcting ship trajectory intelligent prediction method according to claim 1, characterized in that, The setting position prediction error threshold is setting a position prediction error threshold to wherein, , represent the mean and bias of the prediction error of the XGBoost model on the training set.

Images