A track prediction method and device based on a pre-trained large language model

Abstract

This invention provides a method and apparatus for trajectory prediction based on a pre-trained large language model, relating to the field of artificial intelligence. The method includes: acquiring historical trajectory data of a target aircraft and performing data preprocessing to obtain a trajectory sequence; constructing prompt words and inputting the prompt words into a pre-trained large language model to obtain word embedding representations; the prompt words contain a description of the trajectory prediction task and a description of the trajectory data; encoding the trajectory sequence to obtain an encoded trajectory sequence and a trajectory position sequence, and segmenting the trajectory position sequence into words; constructing a spatiotemporal graph of the trajectory sequence based on the encoded trajectory sequence and extracting spatiotemporal dependency feature representations of the trajectory sequence; fusing the word embedding representations and the spatiotemporal dependency feature representations using the pre-trained large language model to generate a trajectory information feature representation; and inputting the trajectory information feature representations into a trained prediction module to obtain predicted values ​​of the longitude, latitude, and altitude of the target aircraft at future times.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a method and apparatus for trajectory prediction based on a pre-trained large language model. Background Technology

[0002] Trajectory Based Operation (TBO) effectively reduces flight trajectory uncertainty, improves airspace resource utilization, and ensures aviation safety by accurately predicting aircraft flight time and trajectory. Trajectory prediction uses historical three-dimensional position and speed information to predict the aircraft's three-dimensional position (longitude, latitude, and altitude) over a future period to assist air traffic control.

[0003] Currently, domestic and international scholars mainly use the following three research methods for trajectory prediction:

[0004] (1) Methods based on aerodynamics or kinematic models. These methods build predictive models based on kinematic assumptions and require numerous parameters such as atmospheric conditions, pilot actions, and flight intentions. Kinematic models are mostly implemented under some ideal assumptions and rarely consider the influence of actual constraints and human behavior. In addition, these methods require a large number of parameters, some of which are quite sensitive and difficult to obtain. Once the data source is limited, the predictive accuracy of the model will be greatly reduced, or even become unusable.

[0005] (2) Methods based on parameterless estimation, such as various improved Kalman filter algorithms. These algorithms require a lot of modeling and analysis work, and their internal structure is relatively simple, making it impossible to fully explore the hidden dependencies in historical data, resulting in insufficient model robustness.

[0006] (3) Deep learning-based methods. With the rapid development of deep learning, researchers have proposed trajectory prediction models based on LSTM, bidirectional LSTM, CNN, Transformer, etc., to mine the changing patterns of aircraft trajectories from a large amount of historical trajectory data. These methods mostly model the trajectory prediction problem as a multivariate time series problem, neglecting the learning and modeling of semantic correlations between trajectory positions, and the prediction accuracy of the models still needs to be improved. Summary of the Invention

[0007] To improve the accuracy of trajectory prediction by introducing learning and modeling of semantic correlations between aircraft trajectory positions, this invention proposes a trajectory prediction method and device based on a pre-trained large language model. By utilizing the text understanding and reasoning capabilities of the existing pre-trained large language model, the semantic correlations between trajectory point positions can be better modeled, thus assisting in trajectory prediction.

[0008] In a first aspect, the present invention provides a trajectory prediction method based on a pre-trained large language model, comprising:

[0009] Step 1: Obtain the historical flight path data of the target aircraft and perform data preprocessing to obtain a flight path sequence; wherein, the information of each track point in the flight path sequence includes longitude, latitude, altitude and speed;

[0010] Step 2: Construct prompt words and input them into a pre-trained large language model to obtain word embedding representations; wherein, the prompt words include a description of the trajectory prediction task and a description of the trajectory data;

[0011] Step 3: Generate geocoding for the longitude, latitude, and altitude of each track point in the track sequence to obtain the corresponding track position sequence. Combine the speed as an attribute of each track point with the track position sequence to obtain the encoded track sequence, and then perform word segmentation on the track position sequence.

[0012] Step 4: Construct a spatiotemporal graph of the track sequence based on the encoded track sequence, and extract the spatiotemporal dependency feature representation of the track sequence from the spatiotemporal graph and the track position sequence;

[0013] Step 5: Use a pre-trained large language model to fuse the word embedding representation and the spatiotemporal dependency feature representation to generate a trajectory information feature representation;

[0014] Step 6: Input the trajectory information feature representation into the trained prediction module to obtain the predicted longitude, latitude and altitude of the target aircraft at future times.

[0015] Furthermore, in step 3, geocoding is generated for the longitude, latitude, and altitude of each track point in the track sequence, specifically including:

[0016] The geohash algorithm is used to encode longitude and latitude separately, generating their respective binary codes;

[0017] Set a range of height values, and continuously divide the height into two parts based on the midpoint of the height range until a binary code with the required precision is obtained.

[0018] Furthermore, in step 3, the BPE word segmenter is used to segment the track position sequence.

[0019] Furthermore, in step 4, a spatiotemporal graph of the track sequence is constructed based on the encoded track sequence, specifically including:

[0020] The flight path sequence is sliced ​​according to a preset time window;

[0021] For each slice, each trackpoint within the slice is used as a node in the graph, and edges are constructed based on the distances between the trackpoints. This connects the aircraft's speed V in the longitude, latitude, and altitude directions. x V y V z As an attribute of each track point, the corresponding slice graph is constructed; where two track points with the same m-bit geohash prefix are regarded as neighboring points and an edge is established.

[0022] Further, in step 4, the trained trajectory spatiotemporal embedding module is used to extract the spatiotemporal dependency feature representation of the trajectory sequence from the spatiotemporal map and the trajectory position sequence; wherein, the trajectory spatiotemporal embedding module includes a feature extraction unit, a first fully connected layer, a multi-head attention layer and a second fully connected layer, wherein, the feature extraction unit includes a first spatiotemporal convolutional network TCN layer, a GCN layer and a second spatiotemporal convolutional network TCN layer connected in sequence;

[0023] The first and second TCN layers are used to extract the temporal correlation features of the spatiotemporal graph in the time dimension; the encoded track sequence and the output of the first TCN layer are used as the input of the GCN layer, and the spatial correlation features of the spatiotemporal graph are extracted in the spatial dimension; the input of the first fully connected layer is the word embedding of the track position sequence.

[0024] The spatiotemporal dependency features output by the feature extraction unit are used as query vectors, and the output of the first fully connected layer is used as key vectors and value vectors. The spatiotemporal dependency features are mapped to the latent semantic space associated with the text using a multi-head attention layer, and then passed through the second fully connected layer to obtain the spatiotemporal dependency feature representation.

[0025] Secondly, the present invention provides a trajectory prediction device based on a pre-trained large language model, comprising:

[0026] The data preprocessing module is used to acquire historical flight path data of the target aircraft and perform data preprocessing to obtain a flight path sequence; wherein, the information of each flight path point in the flight path sequence includes longitude, latitude, altitude and speed;

[0027] The prompt word embedding representation generation module is used to construct prompt words and input the prompt words into a pre-trained large language model to obtain word embedding representations; wherein, the prompt words include a description of the trajectory prediction task and a description of the trajectory data;

[0028] The track sequence encoding and word segmentation module is used to generate geocoding for the longitude, latitude, and altitude of each track point in the track sequence to obtain the corresponding track position sequence. The speed is used as an attribute of each track point and combined with the track position sequence to obtain the encoded track sequence. The track position sequence is then segmented into words.

[0029] The track sequence spatiotemporal graph construction and recoding module is used to construct a spatiotemporal graph of the track sequence based on the encoded track sequence, and extract the spatiotemporal dependency feature representation of the track sequence from the spatiotemporal graph and the track position sequence;

[0030] The trajectory information feature representation module is used to fuse the word embedding representation and the spatiotemporal dependency feature representation using a pre-trained large language model to generate trajectory information feature representation;

[0031] The trajectory prediction module is used to input the trajectory information feature representation into the trained prediction module to obtain the predicted values ​​of the longitude, latitude and altitude of the target aircraft at future times.

[0032] Thirdly, the present invention provides an electronic device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the program, implements the method as described in the first aspect.

[0033] Fourthly, the present invention provides a non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that the computer program, when executed by a processor, implements the method described in the first aspect.

[0034] The beneficial effects of this invention are:

[0035] (1) This invention re-encodes the trajectory data and maps it to the text space. Thus, without fine-tuning the existing pre-trained large language model, only a small number of parameters need to be trained to learn the semantic correlation of trajectory points and realize the trajectory prediction task, providing more reliable technical support for trajectory prediction.

[0036] (2) In order to utilize the semantic correlation of trajectory point locations to provide more accurate trajectory prediction results, this invention constructs and recodes a spatiotemporal graph of the trajectory sequence: First, the longitude, latitude, and altitude of the trajectory are converted from numerical values ​​into textual geocoding. Then, the trajectory sequence is modeled as a spatiotemporal dynamic graph. Temporal convolutional neural networks and graph convolutional neural networks are used to learn the temporal and spatial dependency features of the trajectory spatiotemporal graph. Then, the spatiotemporal dependency features are modally aligned with the text modality through an attention mechanism. The spatiotemporal dependency features of the trajectory points are transformed into a latent semantic space associated with the text, which facilitates the pre-trained large language model to learn and reason about them.

[0037] (3) The encoding of longitude, latitude and altitude designed in this invention converts longitude, latitude and altitude from numerical values ​​into text form, avoiding the problems of model training difficulties and damage to prediction accuracy caused by the large difference between longitude, latitude and altitude units.

[0038] (4) The present invention inputs the trajectory data description and prediction task description through prompt words, which can better guide the large model to reason and also supplement the features of the trajectory sequence to a certain extent. When the re-encoded trajectory sequence is embedded into the vector representation and input together into the pre-trained large language model, the robustness and accuracy of trajectory prediction can be improved. Attached Figure Description

[0039] Figure 1 A flowchart illustrating a trajectory prediction method based on a pre-trained large language model, provided as an embodiment of the present invention;

[0040] Figure 2 A trajectory prediction framework based on a pre-trained large language model is provided as an embodiment of the present invention.

[0041] Figure 3 A structural block diagram of a trajectory prediction device based on a pre-trained large language model provided in an embodiment of the present invention;

[0042] Figure 4 This is a structural block diagram of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0043] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of the embodiments of this invention will be clearly described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0044] Combination Figure 1 and Figure 2 As shown, this embodiment of the invention provides a trajectory prediction method based on a pre-trained large language model, comprising the following steps:

[0045] S101: Acquire historical flight path data of the target aircraft and perform data preprocessing to obtain a flight path sequence; wherein, the information of each flight path point in the flight path sequence includes longitude, latitude, altitude and speed;

[0046] Specifically, the cleaned track sequences are obtained through data preprocessing methods such as data filtering and data standardization. Data filtering mainly includes removing outliers from track points and directly deleting track data with too many missing values. Data standardization mainly includes normalizing the velocity.

[0047] S102: Construct prompt words and input the prompt words into a pre-trained large language model to obtain word embedding representations; wherein, the prompt words include a description of the trajectory prediction task and a description of the trajectory data;

[0048] Specifically, pre-trained large language models can employ LLAMA 3, GLM4, etc. The trajectory prediction task description can be: "Based on the provided aircraft's longitude, latitude, altitude, and velocity sequence in the longitude, latitude, and altitude directions, predict the aircraft's longitude, latitude, and altitude for the next N time steps." The trajectory data description mainly consists of statistical descriptions of a segment of trajectory data, including the range of longitude, latitude, and altitude, the maximum, minimum, and average velocities in each direction, and the time intervals between each trajectory point, among other basic descriptive information.

[0049] After concatenating the prompt words for the trajectory prediction task description and the trajectory data description, the system performs word segmentation using a pre-trained large language model to generate the word embedding representation P = [p1, p2, ..., p...]. n ]∈R n×d , where d is the dimension of the word vector and n is the number of tokens obtained after segmenting the prompt word.

[0050] In addition, flight information, weather information, and other relevant data can be added to the prompts to improve the accuracy of the model's predictions.

[0051] S103: Generate geocoding for the longitude, latitude, and altitude of each track point in the track sequence to obtain the corresponding track position sequence. Combine the speed as an attribute of each track point with the track position sequence to obtain the encoded track sequence, and perform word segmentation on the track position sequence.

[0052] Specifically, longitude and latitude are encoded using the geohash algorithm, converting them into K-bit base32 encoded values. The larger the K value, the higher the accuracy of the geohash value. When K=6, the geohash error is approximately 0.61km. Geohash is a geocoding method, a hierarchical data structure that continuously divides space into grids using a binary method. Geohash is a practical application of the Z-order curve in spatial filling curves. Geohash has a property related to the Z-order curve: geohash strings near a point (but not absolutely) always have a common prefix, and the longer the common prefix, the closer the two points are.

[0053] For height, this embodiment of the invention also refers to the binary search approach of geohash for encoding. To reduce encoding complexity, the height value is first divided by 10, and then a height range [0, max_height] is determined. A binary search is then performed based on the midpoint of the range, max_height / 2, and this process is repeated until a binary code meeting the required precision is obtained. Similarly, for a longitude range of [-180, 180], a point with a longitude of 104.67 is assigned to the right half of the interval and encoded as 1 during the first binary search; conversely, a point with a longitude of less than 104.67 is assigned to the left half and encoded as 0.

[0054] Furthermore, the binary codes corresponding to longitude, latitude, and altitude are converted into base32 codes, that is, 5 bits of binary code are converted into one bit of base32 code.

[0055] It should be noted that the velocity sequences in the three directions of longitude, latitude, and altitude do not need to be encoded or processed. The velocity sequences will be used as attributes of the waypoints to assist the model in learning.

[0056] After encoding and processing the longitude, latitude, and altitude in the flight track sequence, a word segmenter is used to segment the data, representing the flight track position sequence with a sequence of segmented token IDs. This sequence can then be input into a large language model, leveraging the model's text understanding capabilities to learn the semantic relevance between flight track points. The word segmenter can employ BPE (BytePair Encoding), which selects the most frequent adjacent words and merges them into a single token.

[0057] S104: Construct a spatiotemporal graph of the track sequence based on the encoded track sequence, and extract the spatiotemporal dependency feature representation of the track sequence from the spatiotemporal graph and the track position sequence.

[0058] Specifically, for constructing a spatiotemporal graph of a flight path sequence, one possible implementation is to slice the flight path sequence according to a time window T, and construct a graph for the flight path sequence within each slice: each flight path point within each slice is a node of the graph, and edges are constructed based on the distance between the flight path points (edges are established between two flight path points whose distance is less than a threshold). The aircraft's speed V in the longitude, latitude, and altitude directions is also considered. x V y V z As an attribute of each waypoint, the location of a waypoint can be measured using the common prefix characteristics of longitude and latitude encoding. Two waypoints with the same m-bit geohash prefix are considered neighboring points and can be connected; the value of m is determined based on the geohash accuracy.

[0059] For extracting spatiotemporal dependency feature representations of track sequences, such as Figure 2As shown, a trained spatiotemporal embedding module is used to extract spatiotemporal dependency feature representations of the track sequence from the spatiotemporal map and the track position sequence. This spatiotemporal embedding module mainly includes a feature extraction unit, a first fully connected layer, a multi-head attention layer, and a second fully connected layer. The feature extraction unit includes a first spatiotemporal convolutional network (TCN) layer, a second spatiotemporal convolutional network (GCN) layer, and a third spatiotemporal convolutional network (TCN) layer connected sequentially. This embodiment proposes using two different temporal convolutional networks (TCNs) to fully extract the temporal correlation features of the spatiotemporal map in the temporal dimension. The first TCN layer uses a larger convolutional kernel to extract global temporal features of the spatiotemporal map, while the second TCN layer uses a smaller convolutional kernel to extract local temporal features. The encoded track sequence and the output of the first TCN layer are used as inputs to the GCN layer, which extracts spatial correlation features on the slice map of each slice. The input to the first fully connected layer is the word embedding of the track position sequence (this word embedding can be obtained using a pre-trained word vector model, such as the BERT word vector model or existing LLM-generated word embeddings). Then, the spatiotemporal dependency features output by the feature extraction unit are used as the query vector, and the output of the first fully connected layer is used as the key vector and value vector. The obtained spatiotemporal dependency features are mapped to the latent semantic space associated with the text using a multi-head attention mechanism, thereby obtaining a trajectory sequence embedding vector representation after the spatiotemporal graph of the track sequence is aligned with the text modality, which is the final spatiotemporal dependency feature representation required.

[0060] S105: The word embedding representation and the spatiotemporal dependent feature representation are fused using a pre-trained large language model to generate a trajectory information feature representation.

[0061] Specifically, the vectors from steps 102 and 104 are concatenated and input into a pre-trained large language model with frozen parameters to obtain the final feature representation of the trajectory information. Through this step, the description of the trajectory prediction task, the basic statistical information of the trajectory data, and the spatiotemporal dependence features of the trajectory points are fused using the large language model.

[0062] S106: Input the trajectory information feature representation into the trained prediction module to obtain the predicted values ​​of the longitude, latitude and altitude of the target aircraft at future times.

[0063] Specifically, during training, the feature representation of the flight path information from the training set is input into the prediction module. The model parameters are optimized by minimizing the mean squared error loss function to predict the aircraft's longitude, latitude, and altitude at future times. The structure of the prediction module is as follows: Figure 2 As shown.

[0064] The trajectory prediction method provided in this invention, without changing the existing pre-trained large language model, adds a pre-programmed trajectory recoding module and prompt words for guidance. It only requires training a small number of parameters and can use the existing large language model to achieve the trajectory prediction task, providing more reliable technical support for trajectory prediction.

[0065] This invention addresses the geocoding of longitude, latitude, and altitude by converting these numerical values ​​into textual form. This avoids the problems of model training difficulties and reduced prediction accuracy caused by the large differences between longitude, latitude, and altitude units in the past.

[0066] Corresponding to the methods mentioned above, such as Figure 3 As shown, this embodiment of the invention also provides a trajectory prediction device based on a pre-trained large language model, including: a data preprocessing module, a prompt word embedding representation generation module, a trajectory sequence encoding and word segmentation module, a trajectory sequence spatiotemporal graph construction and recoding module, a trajectory information feature representation module, and a trajectory prediction module.

[0067] The system comprises several modules: a data preprocessing module for acquiring historical flight path data of the target aircraft and performing preprocessing to obtain a flight path sequence; where each track point in the flight path sequence includes longitude, latitude, altitude, and speed; a prompt word embedding representation generation module for constructing prompt words and inputting them into a pre-trained large language model to obtain word embedding representations; where the prompt words include a description of the flight path prediction task and a description of the flight path data; a flight path sequence encoding and word segmentation module for generating geocodings for the longitude, latitude, and altitude of each track point in the flight path sequence to obtain the corresponding flight path location sequence, combining speed as an attribute of each track point with the flight path location sequence to obtain the encoded flight path sequence, and performing word segmentation on the flight path location sequence; and a flight path sequence spatiotemporal graph construction and recoding module for constructing a spatiotemporal graph of the flight path sequence based on the encoded flight path sequence, and extracting spatiotemporal dependency feature representations of the flight path sequence from the spatiotemporal graph and the flight path location sequence. The trajectory information feature representation module is used to fuse the word embedding representation and the spatiotemporal dependency feature representation using a pre-trained large language model to generate a trajectory information feature representation. The trajectory prediction module is used to input the trajectory information feature representation into the trained prediction module to obtain the predicted longitude, latitude, and altitude of the target aircraft at future times.

[0068] The apparatus provided in this embodiment of the invention is for implementing the above method. Its specific functions can be found in the above method embodiments, and will not be repeated here.

[0069] In the device provided in this embodiment of the invention, the track sequence spatiotemporal graph construction and recoding module first converts the longitude, latitude, and altitude of the track from numerical values ​​into geocoding in text form. Then, it models the track sequence as a spatiotemporal dynamic graph. Using temporal convolutional neural networks and graph convolutional neural networks, the temporal and spatial dependency features of the track sequence spatiotemporal graph can be effectively captured. Then, through an attention mechanism, the spatiotemporal dependency features are modally aligned with the text modality, and the spatiotemporal dependency features of the track points are transformed into a latent semantic space associated with the text, thereby facilitating the pre-trained large language model to learn and reason about it.

[0070] This invention uses prompt words to input the basic features of track data and the description of the prediction task, thereby guiding a large language model to perform reasoning. It also maps the text and the spatiotemporal graph of the track sequence into a unified semantic space for effective interaction, resulting in high processing efficiency and making it suitable for both long-term and short-term track prediction tasks.

[0071] Figure 4 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 4 As shown, the electronic device may include: a processor 401, a communication interface 402, a memory 403, and a communication bus 404. The processor 401, communication interface 402, and memory 403 communicate with each other via the communication bus 404. The processor 401 can call logical instructions in the memory 403 to execute a trajectory prediction method based on a pre-trained large language model. This method includes: acquiring historical trajectory data of the target aircraft and performing data preprocessing to obtain a trajectory sequence; wherein the information of each trajectory point in the trajectory sequence includes longitude, latitude, altitude, and speed; constructing prompt words and inputting the prompt words into the pre-trained large language model to obtain word embedding representations; wherein the prompt words include a trajectory prediction task description and a trajectory data description; generating geocoding for the longitude, latitude, and altitude of each trajectory point in the trajectory sequence to obtain a geocoding representation for the target aircraft. The corresponding flight path position sequence is used to combine the speed as an attribute of each flight path point with the flight path position sequence to form an encoded flight path sequence, and the flight path position sequence is segmented into words; a spatiotemporal graph of the flight path sequence is constructed based on the encoded flight path sequence, and spatiotemporal dependency feature representations of the flight path sequence are extracted from the spatiotemporal graph and the flight path position sequence; the word embedding representations and the spatiotemporal dependency feature representations are fused using a pre-trained large language model to generate a flight path information feature representation; the flight path information feature representation is input into a trained prediction module to obtain the predicted longitude, latitude and altitude of the target aircraft at future times.

[0072] Furthermore, when the logical instructions in the aforementioned memory 403 are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0073] This invention also provides a computer program product, which includes a computer program stored on a non-transitory computer-readable storage medium. The computer program includes program instructions, and when these instructions are executed by a computer, the computer can perform the methods provided in the above-described method embodiments, such as: acquiring historical flight path data of a target aircraft and performing data preprocessing to obtain a flight path sequence; wherein the information of each flight path point in the flight path sequence includes longitude, latitude, altitude, and speed; constructing prompt words and inputting the prompt words into a pre-trained large language model to obtain a word embedding representation; wherein the prompt words include a description of the flight path prediction task and a description of the flight path data. The system generates geocodings for the longitude, latitude, and altitude of each track point in the track sequence to obtain the corresponding track position sequence. It then combines speed as an attribute of each track point with the track position sequence to form the encoded track sequence, and performs word segmentation on the track position sequence. Based on the encoded track sequence, it constructs a spatiotemporal graph of the track sequence and extracts spatiotemporal dependency feature representations of the track sequence from the spatiotemporal graph and the track position sequence. A pre-trained large language model is used to fuse the word embedding representations and the spatiotemporal dependency feature representations to generate track information feature representations. These track information feature representations are then input into a trained prediction module to obtain predicted values ​​for the longitude, latitude, and altitude of the target aircraft at future times.

[0074] This invention also provides a non-transitory computer-readable storage medium storing a computer program thereon. When executed by a processor, the computer program implements the methods provided in the above-described method embodiments, including, for example,: acquiring historical flight path data of a target aircraft and performing data preprocessing to obtain a flight path sequence; wherein, the information of each flight path point in the flight path sequence includes longitude, latitude, altitude, and speed; constructing prompt words and inputting the prompt words into a pre-trained large language model to obtain word embedding representations; wherein, the prompt words include a description of the flight path prediction task and a description of the flight path data; and processing the longitude, latitude, altitude, and speed of each flight path point in the flight path sequence. The system generates geocoding to obtain the corresponding track position sequence. Velocity is used as an attribute of each track point and combined with the track position sequence to form the encoded track sequence. The track position sequence is then segmented into words. A spatiotemporal graph of the track sequence is constructed based on the encoded track sequence. Spatiotemporal dependency feature representations of the track sequence are extracted from the spatiotemporal graph and the track position sequence. A pre-trained large language model is used to fuse the word embedding representations and the spatiotemporal dependency feature representations to generate track information feature representations. These track information feature representations are then input into a trained prediction module to obtain predicted values ​​for the longitude, latitude, and altitude of the target aircraft at future times.

[0075] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0076] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A trajectory prediction method based on a pre-trained large language model, characterized in that, include: Step 1: Obtain the historical flight path data of the target aircraft and perform data preprocessing to obtain a flight path sequence; wherein, the information of each track point in the flight path sequence includes longitude, latitude, altitude and speed; Step 2: Construct prompt words and input them into a pre-trained large language model to obtain word embedding representations; wherein, the prompt words include a description of the trajectory prediction task and a description of the trajectory data; Step 3: Generate geocoding for the longitude, latitude, and altitude of each track point in the track sequence to obtain the corresponding track position sequence. Combine the speed as an attribute of each track point with the track position sequence to obtain the encoded track sequence, and then perform word segmentation on the track position sequence. Step 4: Construct a spatiotemporal graph of the track sequence based on the encoded track sequence, and extract the spatiotemporal dependency feature representation of the track sequence from the spatiotemporal graph and the track position sequence using a trained track spatiotemporal embedding module; wherein, the track spatiotemporal embedding module includes a feature extraction unit, a first fully connected layer, a multi-head attention layer and a second fully connected layer, wherein, the feature extraction unit includes a first spatiotemporal convolutional network TCN layer, a GCN layer and a second spatiotemporal convolutional network TCN layer connected in sequence; The first and second TCN layers are used to extract the temporal correlation features of the spatiotemporal graph in the time dimension; the encoded track sequence and the output of the first TCN layer are used as the input of the GCN layer, and the spatial correlation features of the spatiotemporal graph are extracted in the spatial dimension; the input of the first fully connected layer is the word embedding of the track position sequence. The spatiotemporal dependent features output by the feature extraction unit are used as query vectors, and the output of the first fully connected layer is used as key vectors and value vectors. The spatiotemporal dependent features are mapped to the latent semantic space associated with the text using a multi-head attention layer, and then passed through the second fully connected layer to obtain the spatiotemporal dependent feature representation. Step 5: Use a pre-trained large language model to fuse the word embedding representation and the spatiotemporal dependency feature representation to generate a trajectory information feature representation; Step 6: Input the trajectory information feature representation into the trained prediction module to obtain the predicted longitude, latitude and altitude of the target aircraft at future times.

2. The trajectory prediction method based on a pre-trained large language model according to claim 1, characterized in that, Step 3 involves generating geocodings for the longitude, latitude, and altitude of each track point in the track sequence, specifically including: The geohash algorithm is used to encode longitude and latitude separately, generating their respective binary codes; Set a range of height values, and continuously divide the height into two parts based on the midpoint of the height range until a binary code with the required precision is obtained.

3. The trajectory prediction method based on a pre-trained large language model according to claim 1, characterized in that, In step 3, the BPE word segmenter is used to segment the track position sequence.

4. The trajectory prediction method based on a pre-trained large language model according to claim 2, characterized in that, In step 4, a spatiotemporal graph of the track sequence is constructed based on the encoded track sequence, specifically including: The flight path sequence is sliced ​​according to a preset time window; For each slice, each trackpoint within the slice is used as a node in the graph, and edges are constructed based on the distances between the trackpoints. This connects the aircraft's speed V in the longitude, latitude, and altitude directions. x V y V z As an attribute of each track point, the corresponding slice graph is constructed; where two track points with the same m-bit geohash prefix are regarded as neighboring points and an edge is established.

5. A trajectory prediction device based on a pre-trained large language model, characterized in that, include: The data preprocessing module is used to acquire historical flight path data of the target aircraft and perform data preprocessing to obtain a flight path sequence; wherein, the information of each flight path point in the flight path sequence includes longitude, latitude, altitude and speed; The prompt word embedding representation generation module is used to construct prompt words and input the prompt words into a pre-trained large language model to obtain word embedding representations; wherein, the prompt words include a description of the trajectory prediction task and a description of the trajectory data; The track sequence encoding and word segmentation module is used to generate geocoding for the longitude, latitude, and altitude of each track point in the track sequence to obtain the corresponding track position sequence. The speed is used as an attribute of each track point and combined with the track position sequence to obtain the encoded track sequence. The track position sequence is then segmented into words. The track sequence spatiotemporal graph construction and recoding module is used to construct a spatiotemporal graph of the track sequence based on the encoded track sequence, and to extract the spatiotemporal dependency feature representation of the track sequence from the spatiotemporal graph and the track position sequence using a trained track spatiotemporal embedding module; wherein, the track spatiotemporal embedding module includes a feature extraction unit, a first fully connected layer, a multi-head attention layer and a second fully connected layer, wherein, the feature extraction unit includes a first spatiotemporal convolutional network TCN layer, a GCN layer and a second spatiotemporal convolutional network TCN layer connected in sequence; The first and second TCN layers are used to extract the temporal correlation features of the spatiotemporal graph in the time dimension; the encoded track sequence and the output of the first TCN layer are used as the input of the GCN layer, and the spatial correlation features of the spatiotemporal graph are extracted in the spatial dimension; the input of the first fully connected layer is the word embedding of the track position sequence. The spatiotemporal dependent features output by the feature extraction unit are used as query vectors, and the output of the first fully connected layer is used as key vectors and value vectors. The spatiotemporal dependent features are mapped to the latent semantic space associated with the text using a multi-head attention layer, and then passed through the second fully connected layer to obtain the spatiotemporal dependent feature representation. The trajectory information feature representation module is used to fuse the word embedding representation and the spatiotemporal dependency feature representation using a pre-trained large language model to generate trajectory information feature representation; The trajectory prediction module is used to input the trajectory information feature representation into the trained prediction module to obtain the predicted values ​​of the longitude, latitude and altitude of the target aircraft at future times.

6. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the method as described in any one of claims 1 to 4.

7. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method as described in any one of claims 1 to 4.

Images