A flight trajectory prediction method and system based on spatial information perception

Abstract

The application discloses a flight trajectory prediction method and system based on space information perception, comprising: obtaining historical aircraft trajectory information, performing multi-scale flight trajectory segmentation based on down-sampling; calculating aircraft flight mode correlation graphs for aircraft trajectory information of different scales; constructing a high-order tensor graph neural network based on the flight mode correlation graphs, and fusing trajectory information of different aircrafts in the same airspace; obtaining historical wind speed and direction information, and fusing the wind speed and direction information with the aircraft trajectory information; fusing the fusion results according to different scales, and outputting a predicted future aircraft trajectory. The application can ensure that the information of different aircrafts in the same airspace is fully interacted, and additional surrounding wind speed and direction information is integrated as auxiliary features, so that the flight trajectory can be accurately predicted.

Description

Technical Field

[0001] This invention belongs to the field of low-altitude flight management technology, specifically relating to a flight trajectory prediction method and system based on airspace information perception. Background Technology

[0002] Flight trajectory prediction, a core task of air traffic management, plays a crucial role in airspace resource optimization, flight scheduling decisions, flight safety monitoring, and conflict detection and resolution. Current mainstream trajectory prediction methods primarily rely on time-series modeling of aircraft state parameters, employing techniques such as autoregressive models and long short-term memory networks (LSTMs), achieving a certain level of accuracy in short-term predictions under simple flight scenarios. However, with the continuous increase in airspace traffic density and the increasing complexity of flight missions, traditional methods are gradually revealing significant limitations. Existing methods often treat aircraft as isolated entities, failing to fully consider key factors such as the dynamic interactions between multiple aircraft within the airspace, airspace structural constraints, and the impact of meteorological conditions. This leads to significant deviations between predicted trajectories and actual flight paths in complex scenarios such as intersecting routes, terminal area arrivals and departures, and air traffic convergence. The prediction accuracy and robustness are insufficient to meet the refined requirements of modern air traffic management.

[0003] In recent years, the development of deep learning methods, especially spatiotemporal graph neural networks, has provided new technical pathways for flight trajectory prediction. These methods, by modeling the spatiotemporal dependencies among multiple aircraft, have improved the synergy of group trajectory prediction to some extent. However, existing deep learning methods still have some shortcomings: First, most methods only consider the relative geometric relationships between aircraft, failing to effectively incorporate the correlation between aircraft trajectories, resulting in a lack of reference for the model's identification of subsequent aircraft states and prediction of future trajectories. Second, existing methods only consider the characteristics of the aircraft itself, without incorporating surrounding wind speed and direction information. However, aircraft flight is affected by the surrounding environment, therefore existing methods need to fully and reasonably model the relevant states of the aircraft.

[0004] Therefore, there is an urgent need for an intelligent flight trajectory prediction method that can deeply understand airspace structure and effectively integrate multi-source conditional information. This method can ensure sufficient information exchange between different aircraft in the same airspace while incorporating additional surrounding wind speed and direction information as auxiliary features, thereby making more accurate predictions of flight trajectories. Summary of the Invention

[0005] To address the shortcomings of the existing technologies, the present invention aims to provide a flight trajectory prediction method and system based on airspace information perception, which can ensure full information exchange between different aircraft in the same airspace while incorporating additional surrounding wind speed and direction information as auxiliary features, thereby making accurate predictions of flight trajectories.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0007] The present invention provides a flight trajectory prediction method based on airspace information perception, comprising the following steps:

[0008] 1) Obtain historical aircraft trajectory information and perform multi-scale flight trajectory segmentation based on downsampling;

[0009] 2) Calculate the correlation map of aircraft flight modes for aircraft trajectory information at different scales;

[0010] 3) Construct a high-order tensor graph neural network based on flight mode correlation graphs to fuse trajectory information of different aircraft in the same airspace;

[0011] 4) Obtain historical wind speed and direction information, and fuse the wind speed and direction information with aircraft trajectory information;

[0012] 5) Based on the different scales obtained in step 1), fuse the fusion results obtained in step 4) and output the predicted future aircraft trajectory.

[0013] Furthermore, step 1) specifically includes: acquiring historical flight trajectory data. ,in For time step size, This represents the number of aircraft currently in the airspace. To reduce the number of dimensions in the flight trajectory information, a downsampling method was used to optimize the flight trajectory data. To perform the segmentation, the expression is as follows:

[0014] ;

[0015] in, Indicates the historical flight path The trajectory sequence obtained after downsampling. , , This represents the total number of downsampling attempts. Indicates to Perform a convolution operation with a kernel of 2 in the time dimension.

[0016] Further, step 2) specifically includes:

[0017] 21) Set two flight trajectory sequences and The best match between the two flight trajectory sequences is calculated as follows:

[0018] ;

[0019] ;

[0020] in, ; Represents a sequence The middle is located in the first The first position The numerical values ​​of each feature, Represents a sequence The middle is located in the first The first position The numerical values ​​of each feature, express The former Subsequence consisting of positions and The former Subsequence consisting of positions The best match Since the lengths of both flight trajectory sequences are... The best matching of the two sequences can be obtained as follows: ;

[0021] 22) Construct an inter-aircraft distance correlation matrix based on the best matching of flight trajectory sequences among different aircraft within the same airspace. ,as follows:

[0022] ;

[0023] in, Representation matrix The Middle Line 1 The values ​​in the column;

[0024] 23) By calculating the correlation between the differences in the flight trajectory sequences, the correlation matrix between the flight trajectories is obtained. ,as follows:

[0025] ;

[0026] ;

[0027] ;

[0028] ;

[0029] ;

[0030] in, Represents aircraft trajectory sequence In the The value at each position Represents aircraft trajectory sequence In the The position and the first The difference between the values ​​at each position, Represents aircraft trajectory sequence In the The value at each position Represents aircraft trajectory sequence In the The position and the first The difference between the values ​​at each position, express Composed of sequences The middle is located in the first The first position The numerical values ​​of each feature, express Composed of sequences The middle is located in the first The first position The numerical values ​​of each feature, express The former Subsequence consisting of positions and The former Subsequence consisting of positions The best match Since the length of the difference between the two flight trajectory sequences is both The best matching of the two sequences can be obtained as follows: , Representation matrix The Middle Line 1 The values ​​in the column;

[0031] 24) Correlation matrix of distances between aircraft Correlation matrix between aircraft trajectories Weighted summation yields the final aircraft flight mode correlation matrix. ,as follows:

[0032] ;

[0033] in, These are learnable weight parameters.

[0034] Furthermore, step 3) specifically includes:

[0035] 31) Based on the correlation matrix of aircraft flight modes Constructing a second-order correlation tensor ,as follows:

[0036] ;

[0037] in, Correlation matrix of aircraft flight modes The degree matrix is ​​as follows:

[0038] ;

[0039] in, Degree matrix No. Line 1 The values ​​in the column;

[0040] 32) Based on the second-order correlation tensor Construct higher-order correlation tensors :

[0041] ;

[0042] in, Indicates in Order correlation tensor In the middle, node The higher-order interaction relationship, that is, in Order correlation tensor In the middle, the first dimension is a node. The sequence number, the second dimension is the node. The sequence number, and so on, the number of... The dimension of the node The numerical value corresponding to the serial number; In the second-order correlation tensor In the middle, the first dimension is a node. The sequence number, the second dimension is the node. The numerical value corresponding to the serial number;

[0043] 33) Through order-up propagation and order-down propagation between different orders, features of different dimensions are propagated to higher-order and lower-order structures, as follows:

[0044] ;

[0045] ;

[0046] in, This represents the first level in a high-order tensor graph neural network. Layer Hidden features of the order, These are the learnable weight matrices for ascending and descending propagation, respectively. These are the matrices representing the degree of preservation of the hidden features of the current order in the previous layer during the update process in both the ascending and descending propagation. It is a non-linear activation function. In the tensor, the first... Perform tensor-matrix contraction multiplication on the first mode;

[0047] 34) Considering information from the same order, higher order, and lower order, the hidden features are simultaneously propagated to higher-order structures, lower-order structures, and the same-order structure, as follows:

[0048] (1);

[0049] In high-order tensor graph neural networks based on correlation graphs, from the first... Layer to the first The hidden features are propagated through layers, and the final high-order tensor graph neural network based on the correlation graph is obtained by stacking multiple hidden layers based on the above formula (1).

[0050] Further, step 4) specifically includes:

[0051] 41) Obtain the output of a high-order tensor graph neural network and airspace wind speed and direction information ,in For time step size, This represents the number of aircraft currently in the airspace. This represents the number of dimensions in the output of a high-order tensor graph neural network. To determine the number of dimensions for wind speed and direction information, an embedding layer of a high-order tensor graph neural network is used to align the dimensions of the wind speed and direction information with the dimensions of the high-order tensor graph neural network output. The expression is as follows:

[0052] ;

[0053] in, This represents wind speed and direction information after dimension alignment. This indicates an embedding operation, where the feature dimension is... of Transformed into feature dimension of ;

[0054] 42) Considered as a query sequence in the cross-attention mechanism , Considered as key sequences in cross-attention mechanisms Sum sequence Wind speed and direction information are fused using a cross-attention mechanism, as follows:

[0055] ;

[0056] in, This is the output obtained after cross-attention calculation; For cross-attention calculation; for Activation function.

[0057] Furthermore, step 5) specifically includes:

[0058] Setting flight trajectory concealment features at different scales Aggregate and calculate the final output ,as follows:

[0059] ;

[0060] in, The weight matrix is ​​a learnable matrix. It is a fully connected layer. Connect the outputs of different fully connected layers. , This represents the size of the prediction time step in the output.

[0061] Furthermore, the present invention also provides a flight trajectory prediction system based on airspace information perception, comprising:

[0062] The trajectory segmentation module is used to acquire historical aircraft trajectory information and perform multi-scale flight trajectory segmentation based on downsampling;

[0063] The calculation module is used to calculate the correlation map of aircraft flight modes based on aircraft trajectory information at different scales.

[0064] The network construction module is used to construct a high-order tensor graph neural network based on the flight mode correlation map, and to fuse the trajectory information of different aircraft in the same airspace.

[0065] The fusion module is used to acquire historical wind speed and direction information and fuse it with aircraft trajectory information.

[0066] The trajectory prediction module fuses the fusion results based on different scales and outputs the predicted future aircraft trajectory.

[0067] The beneficial effects of this invention are:

[0068] (1) The method of the present invention effectively captures multiple spatial information in flight trajectory prediction by simultaneously calculating the spatial correlation between the geographical locations of aircraft and the spatial correlation between the trajectories of aircraft. This solves the problem that previous flight trajectory prediction methods captured less information about surrounding aircraft, thereby achieving more interpretable and more accurate flight trajectory prediction.

[0069] (2) The method of the present invention uses downsampling to generate multi-scale flight trajectory inputs for the model, so that the model can pay attention to trajectory information at different scales at the same time, and learn fine-grained flight trajectory fine-tuning methods and coarse-grained aircraft flight trends, forming a more comprehensive model of flight trajectory to optimize the final prediction effect.

[0070] (3) The method of the present invention utilizes a high-order tensor graph neural network, which can explicitly model the collective behavior and attributes of node clusters, surpassing the dependence of traditional GNN on pairwise relationships. By allowing information to propagate between different orders, it achieves multi-scale fusion in the spatial domain, optimizing the model's perception and capture of spatial correlation.

[0071] (4) The method of the present invention adopts a cross-attention mechanism, which effectively integrates the multimodal features of flight trajectory and surrounding wind speed and direction, and solves the problem of insufficient capture of environmental information in previous flight trajectory prediction. Attached Figure Description

[0072] Figure 1 This is a flowchart of the method of the present invention;

[0073] Figure 2a This is a two-dimensional visualization of the flight trajectory prediction effect using the method of the present invention;

[0074] Figure 2b A two-dimensional visualization of the flight trajectory prediction results using DGCRN;

[0075] Figure 2c A two-dimensional visualization of the flight trajectory prediction results using PDFormer;

[0076] Figure 3a A three-dimensional visualization of the flight trajectory prediction using the method of the present invention;

[0077] Figure 3b A 3D visualization of the flight trajectory prediction using DGCRN;

[0078] Figure 3c A 3D visualization of the flight trajectory prediction using PDFormer. Detailed Implementation

[0079] To facilitate understanding by those skilled in the art, the present invention will be further described below with reference to embodiments and accompanying drawings. The content mentioned in the embodiments is not intended to limit the present invention.

[0080] Reference Figure 1 As shown, a flight trajectory prediction method based on airspace information perception includes the following steps:

[0081] 1) Obtain historical aircraft trajectory information and perform multi-scale flight trajectory segmentation based on downsampling; specifically including: obtaining historical flight trajectory data. ,in For time step size, This represents the number of aircraft currently in the airspace. To reduce the number of dimensions in the flight trajectory information, a downsampling method was used to optimize the flight trajectory data. To perform the segmentation, the expression is as follows:

[0082] ;

[0083] in, Indicates the historical flight path The trajectory sequence obtained after downsampling. , , This represents the total number of downsampling attempts. Indicates to Perform a convolution operation with a kernel of 2 in the time dimension.

[0084] 2) Calculate the correlation map of aircraft flight modes for aircraft trajectory information at different scales; specifically including:

[0085] 21) Set two flight trajectory sequences and The best match between the two flight trajectory sequences is calculated as follows:

[0086] ;

[0087] ;

[0088] in, ; Represents a sequence Located in the middle The first position The numerical values ​​of each feature Represents a sequence Located in the middle The first position The numerical values ​​of each feature express The former Subsequence consisting of positions and The former Subsequence consisting of positions The best match Since the lengths of both flight trajectory sequences are The best matching of the two sequences can be obtained as follows: ;

[0089] 22) Construct an inter-aircraft distance correlation matrix based on the best matching of flight trajectory sequences among different aircraft within the same airspace. ,as follows:

[0090] ;

[0091] in, Representation matrix The Middle Line 1 The values ​​in the column;

[0092] 23) By calculating the correlation between the differences in the flight trajectory sequences, the correlation matrix between the flight trajectories is obtained. ,as follows:

[0093] ;

[0094] ;

[0095] ;

[0096] ;

[0097] ;

[0098] in, Represents aircraft trajectory sequence In the The value at each position Represents aircraft trajectory sequence In the The position and the first The difference between the values ​​at each position, Represents aircraft trajectory sequence In the The value at each position Represents aircraft trajectory sequence In the The position and the first The difference between the values ​​at each position, express Composed of sequences The middle is located in the first The first position The numerical values ​​of each feature, express Composed of sequences The middle is located in the first The first position The numerical values ​​of each feature, express The former Subsequence consisting of positions and The former Subsequence consisting of positions The best match Since the length of the difference between the two flight trajectory sequences is both The best matching of the two sequences can be obtained as follows: , Representation matrix The Middle Line 1 The values ​​in the column;

[0099] 24) Correlation matrix of distances between aircraft Correlation matrix between aircraft trajectories Weighted summation yields the final aircraft flight mode correlation matrix. ,as follows:

[0100] ;

[0101] in, These are learnable weight parameters.

[0102] 3) Construct a high-order tensor graph neural network based on flight pattern correlation maps to fuse trajectory information of different aircraft within the same airspace; specifically including:

[0103] 31) Based on the correlation matrix of aircraft flight modes Constructing a second-order correlation tensor ,as follows:

[0104] ;

[0105] in, Correlation matrix of aircraft flight modes The degree matrix is ​​as follows:

[0106] ;

[0107] in, Degree matrix No. Line 1 The values ​​in the column;

[0108] 32) Based on the second-order correlation tensor Construct higher-order correlation tensors :

[0109] ;

[0110] in, Indicates in Order correlation tensor In the middle, node The higher-order interaction relationship, that is, in Order correlation tensor In the middle, the first dimension is a node. The sequence number, the second dimension is the node. The sequence number, and so on, the number of... The dimension of the node The numerical value corresponding to the serial number; In the second-order correlation tensor In the middle, the first dimension is a node. The sequence number, the second dimension is the node. The numerical value corresponding to the serial number;

[0111] 33) Through order-up propagation and order-down propagation between different orders, features of different dimensions are propagated to higher-order and lower-order structures, as follows:

[0112] ;

[0113] ;

[0114] in, This represents the first level in a high-order tensor graph neural network. Layer Hidden features of the order, These are the learnable weight matrices for ascending and descending propagation, respectively. These are the matrices representing the degree of preservation of the hidden features of the current order in the previous layer during the update process in both the ascending and descending propagation. It is a non-linear activation function. In the tensor, the first... Perform tensor-matrix contraction multiplication on the first mode;

[0115] 34) Considering information from the same order, higher order, and lower order, the hidden features are simultaneously propagated to higher-order structures, lower-order structures, and the same-order structure, as follows:

[0116] (1);

[0117] In high-order tensor graph neural networks based on correlation graphs, from the first... Layer to the first The hidden features are propagated through layers, and the final high-order tensor graph neural network based on the correlation graph is obtained by stacking multiple hidden layers based on the above formula (1).

[0118] 4) Obtain historical wind speed and direction information, and fuse this information with aircraft trajectory information; specifically including:

[0119] 41) Obtain the output of a high-order tensor graph neural network and airspace wind speed and direction information ,in For time step size, This represents the number of aircraft currently in the airspace. This represents the number of dimensions in the output of a high-order tensor graph neural network. To determine the number of dimensions for wind speed and direction information, an embedding layer of a high-order tensor graph neural network is used to align the dimensions of the wind speed and direction information with the dimensions of the high-order tensor graph neural network output. The expression is as follows:

[0120] ;

[0121] in, This represents wind speed and direction information after dimensional alignment. This indicates an embedding operation, where the feature dimension is... of Transformed into feature dimension of ;

[0122] 42) Considered as a query sequence in the cross-attention mechanism , Considered as key sequences in cross-attention mechanisms Sum sequence Wind speed and direction information are fused using a cross-attention mechanism, as follows:

[0123] ;

[0124] in, This is the output obtained after cross-attention calculation; For cross-attention calculation; for Activation function.

[0125] 5) Fuse the fusion results obtained in step 4) based on the different scales obtained in step 1) to output the predicted future aircraft trajectory; specifically including:

[0126] Setting flight trajectory concealment features at different scales Aggregate and calculate the final output ,as follows:

[0127] ;

[0128] in, The weight matrix is ​​a learnable matrix. It is a fully connected layer. Connect the outputs of different fully connected layers. , This represents the size of the prediction time step in the output.

[0129] In the example, the experimental conditions were as follows: all experiments were implemented using PyTorch on an NVIDIA RTX A6000 48G GPU.

[0130] Experiments were conducted on the TrajAir flight trajectory prediction dataset, which contains 111 days of historical aircraft trajectories and wind speed, direction, and environmental background data in the non-tower control terminal airspace surrounding Pittsburgh-Butler regional airports. It includes ADS-B transponder data and corresponding METAR meteorological data. The basic experimental setup integrates the 111 days of flight trajectory information into four 7-day flight trajectory prediction datasets, named "7days1", "7days2", "7days3", and "7days4", respectively. Evaluation metrics include mean squared error (MSE), mean absolute error (MAE), mean displacement error (ADE), and final displacement error (FDE).

[0131] Table 1 shows the comparison results of model trajectory prediction. As shown in Table 1, the model's flight trajectory prediction on this dataset is better than other flight trajectory prediction baseline models.

[0132] Table 1

[0133]

[0134] like Figures 2a-2c As shown, the visualizations on the X and Y axes represent the prediction results of the proposed method, DGCRN, and PDFormer for a flight trajectory in the 7days4 dataset. The figures demonstrate that the proposed method is more accurate in predicting flight trajectories compared to the other two baseline methods. When the trajectory requires the aircraft to turn, it can provide a reasonable prediction that aligns with the actual trajectory. While the other two baseline models also predict a turn, DGCRN's prediction in the final step deviates from the actual physical laws, and PDFormer's predicted final point is significantly different from the actual final point. This verifies that the proposed method has greater rationality and higher accuracy.

[0135] like Figures 3a-3cAs shown, line graphs on three-dimensional coordinates are displayed for the prediction results of a flight trajectory in the 7days4 dataset by the method of this invention, DGCRN, and PDFormer, respectively. It can be seen from the figure that the method of this invention is more accurate in predicting the three-dimensional features of the aircraft than the other two methods. It is not only more accurate in predicting the X-axis and Y-axis coordinates than the other two baseline models, but also makes a better prediction of the slow descent of the aircraft on the Z-axis. This verifies that the method of this invention has higher accuracy in flight trajectory prediction.

[0136] The present invention also provides a flight trajectory prediction system based on airspace information perception, comprising:

[0137] The trajectory segmentation module is used to acquire historical aircraft trajectory information and perform multi-scale flight trajectory segmentation based on downsampling;

[0138] The calculation module is used to calculate the correlation map of aircraft flight modes based on aircraft trajectory information at different scales.

[0139] The network construction module is used to construct a high-order tensor graph neural network based on the flight mode correlation map, and to fuse the trajectory information of different aircraft in the same airspace.

[0140] The fusion module is used to acquire historical wind speed and direction information and fuse it with aircraft trajectory information.

[0141] The trajectory prediction module fuses the fusion results based on different scales and outputs the predicted future aircraft trajectory.

[0142] This invention has many specific applications. The above description is only a preferred embodiment of this invention. It should be noted that for those skilled in the art, several improvements can be made without departing from the principle of this invention, and these improvements should also be considered within the scope of protection of this invention.

Claims

1. A flight trajectory prediction method based on airspace information perception, characterized in that, The steps are as follows: 1) Obtain historical aircraft trajectory information and perform multi-scale flight trajectory segmentation based on downsampling; 2) Calculate the correlation map of aircraft flight modes for aircraft trajectory information at different scales; Step 2) specifically includes: 21) Set two flight trajectory sequences and The best match between the two flight trajectory sequences is calculated as follows: ； ； in, ; Represents a sequence The middle is located in the first The first position The numerical values ​​of each feature Represents a sequence The middle is located in the first The first position The numerical values ​​of each feature express The former Subsequence consisting of positions and The former Subsequence consisting of positions The best match Since the lengths of both flight trajectory sequences are The best matching of the two sequences can be obtained as follows: ; 22) Construct an inter-aircraft distance correlation matrix based on the best matching of flight trajectory sequences among different aircraft within the same airspace. ,as follows: ； in, Representation matrix The Middle Line 1 The values ​​in the column; 23) By calculating the correlation between the differences in the flight trajectory sequences, the correlation matrix between the flight trajectories is obtained. ,as follows: ； ； ； ； ； in, Represents aircraft trajectory sequence In the The value at each position Represents aircraft trajectory sequence In the The position and the first The difference between the values ​​at each position, Represents aircraft trajectory sequence In the The value at each position Represents aircraft trajectory sequence In the The position and the first The difference between the values ​​at each position, express Composed of sequences The middle is located in the first The first position The numerical values ​​of each feature express Composed of sequences The middle is located in the first The first position The numerical values ​​of each feature express The former Subsequence consisting of positions and The former Subsequence consisting of positions The best match Since the length of the difference between the two flight trajectory sequences is both The best matching of the two sequences can be obtained as follows: , Representation matrix The Middle Line 1 The values ​​in the column; 24) Correlation matrix of distances between aircraft Correlation matrix between aircraft trajectories Weighted summation yields the final aircraft flight mode correlation matrix. ,as follows: ； in, These are learnable weight parameters; 3) Construct a high-order tensor graph neural network based on the flight pattern correlation map to fuse trajectory information of different aircraft within the same airspace; Step 3) specifically includes: 31) Based on the correlation matrix of aircraft flight modes Constructing a second-order correlation tensor ,as follows: ； in, Correlation matrix of aircraft flight modes The degree matrix is ​​as follows: ； in, Degree matrix No. Line 1 The values ​​in the column; 32) Based on the second-order correlation tensor Construct higher-order correlation tensors : ； in, Indicates in Order correlation tensor In the middle, node The higher-order interaction relationship, that is, in Order correlation tensor In the middle, the first dimension is a node. The sequence number, the second dimension is the node. The sequence number, and so on, the number of... The dimension position is the node The numerical value corresponding to the serial number; In the second-order correlation tensor In the middle, the first dimension is a node. The sequence number, the second dimension is the node. The numerical value corresponding to the serial number; 33) Through order-up propagation and order-down propagation between different orders, features of different dimensions are propagated to higher-order and lower-order structures, as follows: ； ； in, This represents the first level in a high-order tensor graph neural network. Layer Hidden features of the order, These are the learnable weight matrices for ascending and descending propagation, respectively. These are the matrices representing the degree of preservation of the hidden features of the current order in the previous layer during the update process in both the ascending and descending propagation. It is a non-linear activation function. In the tensor, the first... Perform tensor-matrix contraction multiplication on the first mode; 34) Considering information from the same order, higher order, and lower order, the hidden features are simultaneously propagated to higher-order structures, lower-order structures, and the same-order structure, as follows: （1）； In high-order tensor graph neural networks based on correlation graphs, from the first... Layer to the first The hidden features are propagated through layers, and the final high-order tensor graph neural network based on the correlation graph is obtained by stacking multiple hidden layers based on the above formula (1); 4) Obtain historical wind speed and direction information, and fuse the wind speed and direction information with aircraft trajectory information; 5) Based on the different scales obtained in step 1), fuse the fusion results obtained in step 4) and output the predicted future aircraft trajectory.

2. The flight trajectory prediction method based on airspace information perception according to claim 1, characterized in that, Step 1) specifically includes: acquiring historical flight trajectory data. ,in For time step size, This represents the number of aircraft currently in the airspace. To reduce the number of dimensions in the flight trajectory information, a downsampling method was used to optimize the flight trajectory data. The segmentation is performed using the following expression: ； in, Indicates the historical flight path The trajectory sequence obtained after downsampling. , , This indicates the total number of downsampling attempts. Indicates to Perform a convolution operation with a kernel of 2 in the time dimension.

3. The flight trajectory prediction method based on airspace information perception according to claim 2, characterized in that, Step 4) specifically includes: 41) Obtain the output of a high-order tensor graph neural network and airspace wind speed and direction information ,in For time step size, This represents the number of aircraft currently in the airspace. This represents the number of dimensions in the output of a high-order tensor graph neural network. To determine the number of dimensions for wind speed and direction information, an embedding layer of a high-order tensor graph neural network is used to align the dimensions of the wind speed and direction information with the dimensions of the high-order tensor graph neural network output. The expression is as follows: ； in, This represents wind speed and direction information after dimension alignment. This indicates an embedding operation, where the feature dimension is... of Transformed into feature dimension of ; 42) will Considered as a query sequence in the cross-attention mechanism , Considered as key sequences in cross-attention mechanisms Sum sequence Wind speed and direction information are fused using a cross-attention mechanism, as follows: ； in, This is the output obtained after cross-attention calculation; For cross-attention calculation; for Activation function.

4. The flight trajectory prediction method based on airspace information perception according to claim 3, characterized in that, Step 5) specifically includes: Setting flight trajectory concealment features at different scales Aggregate and calculate the final output ,as follows: ； in, The weight matrix is ​​a learnable matrix. It is a fully connected layer. Connect the outputs of different fully connected layers. , This represents the size of the prediction time step in the output.

5. A flight trajectory prediction system based on airspace information perception, characterized in that, The system is used to implement the flight trajectory prediction method as described in claim 1, and the system includes: The trajectory segmentation module is used to acquire historical aircraft trajectory information and perform multi-scale flight trajectory segmentation based on downsampling; The calculation module is used to calculate the correlation map of aircraft flight modes based on aircraft trajectory information at different scales. The network construction module is used to construct a high-order tensor graph neural network based on the flight mode correlation map, and to fuse the trajectory information of different aircraft in the same airspace. The fusion module is used to acquire historical wind speed and direction information and fuse it with aircraft trajectory information. The trajectory prediction module fuses the fusion results based on different scales and outputs the predicted future aircraft trajectory.

Images