Track prediction method, electronic equipment, storage medium and program product

Abstract

The invention provides a track prediction method, electronic equipment, a storage medium and a program product. The method comprises the following steps: acquiring historical track data of flight equipment in an approaching control area within a preset historical time period; performing time feature extraction on the historical track data through the GRU to obtain time features; the time feature is used for representing a dynamic evolution rule of the historical track data along with time; performing spatial feature extraction on the historical track data through TCN to obtain spatial features; the spatial features are used for representing the time relevance of the spatial state of the flight equipment; processing the time features and the space features through a multi-head attention mechanism module to obtain space-time perception features; and processing the space-time perception characteristics through a decoder to obtain flight path data of the flight equipment in a future preset duration. According to the method, the weights of the hidden features can be adaptively calculated, the differentiation degree of various spatio-temporal features to prediction accuracy is accurately and quantitatively evaluated, and the accuracy of track prediction is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of track prediction, in particular to a track prediction method, an electronic device, a storage medium and a program product. BACKGROUND

[0002] The approach control area covers the key flight segments in the process of flight landing and taking off. In air traffic management (ATM), the traditional flight approach management is based on the judgment of waypoints and sector operation, and lacks accurate track prediction of future flight tracks.

[0003] Currently, 4D track prediction mainly relies on flight performance models, state estimation theory and machine learning methods. The flight performance model predicts the aircraft trajectory based on aircraft intent, state and performance parameters. This prediction method based on aircraft characteristics often faces the problem of parameter estimation difficulty, resulting in inaccurate track prediction. SUMMARY

[0004] The purpose of the embodiments of the present application is to provide a track prediction method, an electronic device, a storage medium and a program product to improve the accuracy of track prediction.

[0005] In a first aspect, the embodiments of the present application provide a track prediction method applied to a track prediction model, the track prediction model being a Transformer architecture model; the track prediction model comprising an encoder and a decoder; the encoder comprising an integrated gated recurrent unit (GRU), a temporal convolutional network (TCN) and a multi-head attention mechanism module; comprising: Obtaining historical track data of a flight device in an approach control area within a preset historical time period; extracting time features of the historical track data by the GRU to obtain time features; the time features are used to represent the dynamic evolution law of the historical track data over time; extracting spatial features of the historical track data by the TCN to obtain spatial features; the spatial features are used to represent the time correlation of the spatial state of the flight device; processing the time features and the spatial features by the multi-head attention mechanism module to obtain spatio-temporal perception features; processing the spatio-temporal perception features by the decoder to obtain track data of the flight device in a future preset time period.

[0006] The embodiments of the present application capture the time dependence of the historical track data through the GRU module, focus on the spatial feature learning of multi-dimensional data through the TCN module, and through the introduction of the multi-head attention mechanism, can adaptively calculate the weight of the hidden features, and according to this, accurately quantify and evaluate the differentiated degree of the prediction accuracy of various spatio-temporal features, thereby improving the accuracy of track prediction.

[0007] In a possible implementation of the first aspect, the GRU includes an update gate and a reset gate; the time feature of the historical track data is extracted through the GRU, and the time feature is obtained, including: inputting the sequence of the historical track data in a time step order into the GRU; for each time step, combining the historical track data corresponding to the current time step with the hidden state of the previous time step maintained in the GRU, calculating the activation value of the update gate and the reset gate in the GRU; and extracting the feature of the historical track data based on the activation value to obtain the time feature.

[0008] In the embodiment of the application, the fusion weight of the historical track data and the current input is dynamically adjusted through the update gate and the reset gate, and then the selection of the gating information is realized, that is, when the track state changes obviously, the update gate can automatically weaken the weight of the historical track data to enhance the capture of the sudden state feature; when the track is in a relatively stable flight state, the reset gate tends to save the long-term historical mode, and when a sudden disturbance is encountered, the reset gate is reduced to suppress the interference of irrelevant historical information, so that the accurate time feature is extracted.

[0009] In a possible implementation of the first aspect, the TCN includes an expanded causal convolution module and a residual module; the spatial feature of the historical track data is extracted through the TCN, and the spatial feature is obtained, including: inputting the historical track data into the expanded causal convolution module, and calculating and obtaining the intermediate feature based on the expansion factor and the convolution kernel size in the expanded causal convolution module; inputting the intermediate feature into the residual module to obtain the spatial feature output by the residual module.

[0010] In the embodiment of the application, the spatial feature is extracted through the TCN module, the long-range dependency modeling capability of the expanded causal convolution is used to associate the track data at different times, and the spatial correlation relationship between multi-dimensional parameters is captured, thereby providing reliable spatial features for subsequent track prediction.

[0011] In a possible implementation of the first aspect, the decoder includes a masked multi-head attention mechanism module and a fully connected layer; the spatiotemporal perception feature is processed through the decoder to obtain the track data of the flight equipment in a future preset time length, including: inputting the spatiotemporal perception feature into the masked multi-head attention mechanism module to obtain a representation sequence output by the masked multi-head attention mechanism module; inputting the representation sequence into the fully connected layer to obtain track data output by the fully connected layer.

[0012] In the embodiment of the application, since the accurate spatiotemporal perception feature is extracted, the accuracy of the track prediction is improved.

[0013] In a possible implementation manner of the first aspect, the method further includes: obtaining training track data in an approach control area, the training track data including longitude, latitude, speed, height, heading, and time of distance zero; constructing the training track data into a plurality of sample target pairs according to time sequence; inputting samples in the sample target pairs into a track prediction model to be trained, and obtaining a prediction result output by the track prediction model to be trained; optimizing parameters of the track prediction model to be trained according to the prediction result, and obtaining a trained track prediction model.

[0014] The embodiments of the present application train the model through the training track data, and the trained track prediction model can accurately predict the track in a future period of time.

[0015] In a possible implementation manner of the first aspect, the method further includes: In the training process, calculating a root mean square error and / or a mean absolute error according to the prediction result and the target; judging a convergence state of the track prediction model in the training process according to the root mean square error and / or the mean absolute error; determining whether to stop the training according to the convergence state.

[0016] In the embodiments of the present application, the root mean square error is more sensitive to large errors, and the mean absolute error can better reflect the actual situation of the prediction value error, and therefore, the effect of the model can be accurately evaluated through the root mean square error and / or the mean absolute error.

[0017] In a possible implementation manner of the first aspect, after the training track data in the approach control area is obtained, the method further includes: performing normalization processing on the training track data by using a Min-Max algorithm.

[0018] The embodiments of the present application perform normalization processing on the historical track data by using the Min-Max algorithm, so that the data is within a certain range, thereby better adapting to the training of the model.

[0019] In a second aspect, the embodiments of the present application provide a track prediction device applied to a track prediction model, the track prediction model being a Transformer architecture model; the track prediction model including a decoder and an encoder; the decoder including an integrated gated recurrent unit GRU, a time convolutional network TCN, and a multi-head attention mechanism module; and the device including: a data acquisition module configured to acquire historical track data of a flight device in an approach control area in a preset historical time period; The time feature extraction module is used to extract time features from historical track data using GRU to obtain time features; the time features are used to characterize the dynamic evolution of historical track data over time. The spatial feature extraction module is used to extract spatial features from historical track data through TCN to obtain spatial features; spatial features are used to characterize the temporal correlation of the spatial state of flight equipment. The feature integration module is used to process temporal and spatial features through the multi-head attention mechanism module to obtain spatiotemporal awareness features; The prediction module is used to process spatiotemporal perception features through a decoder to obtain flight path data of the device for a future preset duration.

[0020] Thirdly, embodiments of this application provide an electronic device, including: a processor, a memory, and a bus, wherein: The processor and memory communicate with each other via a bus; The memory stores program instructions that can be executed by the processor, and the processor can execute the method of the first aspect by calling the program instructions.

[0021] Fourthly, embodiments of this application provide a non-transitory computer-readable storage medium, comprising: A non-transitory computer-readable storage medium stores computer instructions that cause the computer to perform the methods in the various possible implementations of the first aspect.

[0022] Fifthly, embodiments of this application provide a computer program product, including computer program instructions, which, when read and executed by a processor, perform the methods in various possible implementations of the first aspect.

[0023] Other features and advantages of this application will be set forth in the following description and will be apparent in part from the description or may be learned by practicing embodiments of this application. The objectives and other advantages of this application may be realized and obtained by means of the structures particularly pointed out in the written description, claims, and drawings. Attached Figure Description

[0024] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments of this application will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0025] Figure 1 This is a schematic flowchart of a trajectory prediction method provided in an embodiment of this application; Figure 2 A schematic diagram of a trajectory prediction model architecture provided in an embodiment of this application; Figure 3 A schematic diagram of a multi-head attention mechanism module structure provided in an embodiment of this application; Figure 4 A schematic diagram of a GRU structure provided in an embodiment of this application; Figure 5 This application provides a schematic diagram of the structure of an expanded causal convolution module in a TCN. Figure 6 This is a schematic diagram of the residual module structure provided in an embodiment of this application; Figure 7 A schematic diagram of the loss iteration of a trajectory prediction model provided in an embodiment of this application; Figure 8 This is a first model prediction result provided in the embodiments of this application; Figure 9 A schematic diagram illustrating the predictive capability of the first step length provided in an embodiment of this application; Figure 10 A schematic diagram illustrating the predictive capability of the second step length provided in an embodiment of this application; Figure 11 A schematic diagram comparing the RMSE values ​​of different models provided in the embodiments of this application; Figure 12 A schematic diagram comparing the MAE values ​​of different models provided in the embodiments of this application; Figure 13 This is a schematic diagram of the structure of a trajectory prediction device provided in an embodiment of this application; Figure 14 This is a schematic diagram of the physical structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0026] The embodiments of the technical solution of this application will now be described in detail with reference to the accompanying drawings. These embodiments are only used to more clearly illustrate the technical solution of this application and are therefore merely examples, and should not be used to limit the scope of protection of this application.

[0027] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains; the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the application; the terms “comprising” and “having”, and any variations thereof, in the specification, claims, and foregoing description of the drawings are intended to cover non-exclusive inclusion.

[0028] In the description of the embodiments of this application, technical terms such as "first" and "second" are used only to distinguish different objects and should not be construed as indicating or implying relative importance or implicitly specifying the number, specific order, or primary and secondary relationship of the indicated technical features. In the description of the embodiments of this application, "multiple" means two or more, unless otherwise explicitly defined.

[0029] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0030] In the description of the embodiments in this application, the term "and / or" is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, and B existing alone. Additionally, the character " / " in this document generally indicates that the preceding and following related objects have an "or" relationship.

[0031] In the description of the embodiments of this application, the term "multiple" refers to two or more (including two), similarly, "multiple sets" refers to two or more (including two sets), and "multiple pieces" refers to two or more (including two pieces).

[0032] In the description of the embodiments of this application, unless otherwise expressly specified and limited, technical terms such as "installation," "connection," "joining," and "fixing" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral part; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal communication of two components or the interaction between two components. For those skilled in the art, the specific meaning of the above terms in the embodiments of this application can be understood according to the specific circumstances.

[0033] Approach control areas encompass the critical segments of flight landings and takeoffs. In traditional air traffic management (ATM), flight approach control relies on waypoints and sector operations, lacking accurate trajectory prediction. This makes it prone to errors in prediction due to disturbances from other airspace users or complex weather conditions, leading to significant reductions in operational efficiency and safety. To address this, the concept of Trajectory-Based Operation (TBO) is proposed, which solves the problem of controlling large-scale, high-density aircraft within limited airspace under constrained flight intervals. Trajectory prediction is fundamental to TBO, and accurate prediction can advance the development of intelligent air traffic control. By accurately predicting flight trajectories within the approach control area, air traffic management systems can better coordinate flight movements, plan landing sequences, improve departure and arrival efficiency, and achieve more efficient traffic management.

[0034] Currently, trajectory prediction primarily relies on flight performance models, state estimation theory, and machine learning methods. Flight performance models predict aircraft trajectories based on aircraft intent, state, and performance parameters, including climb and descent trajectory models, discrete dynamics models, and trajectory prediction methods that update aircraft intent online. This approach is based on aircraft characteristics but often faces challenges in parameter estimation and low prediction accuracy. State estimation theory focuses on the dynamic relationship between future and past positions without delving into the details of aircraft performance. Typical modeling methods include Kalman filtering and particle filtering, and their improved algorithms.

[0035] In contrast, machine learning methods applied to trajectory prediction are treated as time series prediction tasks, using historical data for training to discover patterns. Classic machine learning models have been applied to trajectory prediction research in various scenarios, but they generally have the following limitations: First, high-dimensional information in the prediction results is often weakened or missing; second, the trajectory generation process lacks consideration of the inherent uncertainty of the data; and third, there is insufficient understanding of the complex characteristics of approach control areas, as approach control areas emphasize structured scenarios such as orderly aircraft approach and dynamic runway allocation, making it difficult to handle long-term trajectory dependencies. Therefore, the above methods have limited adaptability to the special scenarios of approach control areas.

[0036] To address the aforementioned technical challenges, this application provides a trajectory prediction method. This method, based on an encoder-decoder framework and incorporating spatiotemporal feature extraction techniques, constructs a spatiotemporal recognition Transformer trajectory prediction model that combines an Integrated Gated Recurrent Unit (GRU), a Temporal Convolutional Network (TCN), and a multi-head attention mechanism. This model captures the temporal dependencies of trajectory data through the GRU module, while the TCN module focuses on learning the spatial features of multidimensional data. By introducing the multi-head attention mechanism, the model can adaptively calculate the weights of hidden features and accurately quantify the differences in prediction accuracy caused by various spatiotemporal features. Finally, a multi-step recursive prediction mechanism is employed to achieve high-precision continuous prediction of aircraft trajectories. This model emphasizes the complexity of approach control area features and further improves trajectory prediction accuracy compared to traditional models, providing a more effective tool for air traffic management.

[0037] Figure 1 This application provides a schematic flowchart of a trajectory prediction method, which includes: Step 101: Obtain historical flight track data of the flight equipment within the approach control area during a preset historical time period; Step 102: Extract time features from historical track data using GRU to obtain time features; time features are used to characterize the dynamic evolution of historical track data over time. Step 103: Extract spatial features from historical flight track data using TCN to obtain spatial features; spatial features are used to characterize the temporal correlation of the spatial state of flight equipment; Step 104: Process the temporal and spatial features through the multi-head attention mechanism module to obtain spatiotemporal perception features; Step 105: Process the spatiotemporal perception features using a decoder to obtain the flight path data of the device for the future preset duration of flight.

[0038] In step 101, the flight equipment can specifically be an aircraft. Since the embodiments of this application aim to infer the flight status of the flight equipment within a future time period based on its historical state observation sequence, a fixed-length sequential window can be used to achieve this. For example, historical track data within the approach control area within the most recent preset historical time period can be collected, or a preset number of historical track data within the approach control area can be collected within the most recent time period. It should be noted that since the track data is generated based on a fixed time interval, the fixed time interval and preset number can also be converted into a preset historical time period.

[0039] Historical flight track data includes the longitude, latitude, speed, altitude, heading, and UTC (time from zero) of the flight equipment. It may also include information such as the aircraft model and serial number.

[0040] In steps 102 and 103, since there is interdependent spatial information among the historical trajectory data, and the spatial coordinate information contained in the trajectory data is highly coupled with its corresponding temporal evolution process, that is, there is a strong correlation between the spatial and temporal dimensions, the trajectory prediction model provided in this application embodiment is a model based on the Transformer architecture, and on the basis of the Transformer architecture, a deep learning framework with spatiotemporal awareness is constructed by deeply mining multi-dimensional spatiotemporal correlation features. Figure 2 This is a schematic diagram of a trajectory prediction model architecture provided in an embodiment of this application, such as... Figure 2 As shown, the trajectory prediction model includes an encoder and a decoder. The encoder includes an integrated gated recurrent unit (GRU), a temporal convolutional network (TCN), and a multi-head attention mechanism module. The decoder includes a masked multi-head attention mechanism module and a feedforward neural network, etc. The traditional Transformer trajectory prediction model structure includes an input layer, local and global feature extraction layers, and an output layer. The local feature extraction layer consists of ordinary stacked convolutions, while the global feature extraction layer is implemented by the Transformer encoder.

[0041] This application optimizes the traditional model structure by employing dual-channel feature extraction. Temporally, a GRU is used to construct a temporal feature encoder, learning the temporal dynamic dependencies between waypoints through the GRU gating mechanism. Spatially, a dilated causal convolutional network (TCN) is used to construct a temporal convolutional network, learning the spatial relationships between multi-dimensional waypoint parameters through dilation. The TCN joint encoding layer achieves collaborative extraction of spatiotemporal hidden features, ultimately forming a spatiotemporally aware feature vector that is input to the Transformer encoder, enhancing the model's ability to analyze complex spatiotemporally related features.

[0042] It should be noted that the encoder in the trajectory prediction model of this application embodiment may also include a feedforward neural network, a residual module, a layer normalization module, etc. The decoder may also include a feedforward neural network and a residual module. Furthermore, the trajectory prediction model may also include an input layer and an output layer, wherein the input layer is located before the encoder and includes an embedding layer and a position encoding layer. The output layer is located after the decoder and includes a linear layer and a softmax layer.

[0043] To learn temporal patterns from historical flight track data, the GRU (Generative Ratio Requirement) in the flight track prediction model can be used to extract temporal features from the historical flight track data. It can be understood that historical flight track data is actually a spatiotemporal data stream with a clear temporal dependency, and the temporal features are the dynamic evolution patterns of flight state parameters over time.

[0044] TCN is a type of convolutional neural network. Its core advantage lies in the synergistic effect of Dilated Causal Convolution (DCC) and Residual Block (RB) to achieve in-depth mining of multi-dimensional spatial correlation features in historical flight track data. This application embodiment uses cascaded TCN modules to extract spatial features from historical flight track data, which can be used to characterize the temporal correlation of the spatial state of flight equipment.

[0045] In step 104, after obtaining the temporal and spatial features, they are input into the multi-head attention mechanism module. The multi-head attention mechanism allows the model to use different attention weights to model the input sequence. Figure 3 This application provides a schematic diagram of a multi-head attention mechanism module structure, as shown in the embodiments. Figure 3 As shown. The input sequences undergo three separate linear transformations using their respective unique weight matrices: linear transformations of the query, key, and value. The formulas for these linear transformations are as follows:

[0046] in, It is the output of the previous layer, and , and These are the corresponding weight matrices. For a single attention head i, the attention mechanism is calculated as follows: The dot product similarity between the query matrix Q and the key matrix K is calculated. The similarity is then converted into attention weights by applying the softmax function, as shown in the following formula:

[0047] in, This is the dimension of the key; the attention weights are obtained by calculating the softmax function using dot product similarity. Then, the value matrix... We performed a weighted summation to obtain the attention output. .in, .

[0048] They are spliced ​​together and subjected to a linear transformation The spatiotemporal perception features of the final multi-head attention output are obtained. .in, It's about the number of heads that attract attention. It is the weight matrix of the linear transformation.

[0049] It should be noted that the encoder may also include two fully connected layers, with GELU applied as an activation function between the two fully connected layers to perform non-linear mapping and feature extraction on the sub-layer output. The first fully connected layer is as follows: . This is the output of the first fully connected layer. This is the weight matrix. This is a bias term.

[0050] To mitigate the vanishing gradient problem while preserving nonlinear characteristics, the GELU activation function is introduced. When the input is large, the GELU output is approximately linear, which is beneficial for gradient propagation and training in deep networks. Furthermore, its shape resembles a normal distribution, which is advantageous for data modeling. The following equation represents a nonlinear transformation of the output of the first fully connected layer: . This is the output after passing through the GELU activation function.

[0051] Second fully connected layer: . This is the final output of the GRU. This is the weight matrix of the second fully connected layer. This is a bias term.

[0052] Based on the aforementioned multi-head attention mechanism module, spatiotemporal awareness features are obtained through the following method: First, the acquired temporal and spatial features are initially fused, usually by splicing, to form a unified feature representation that combines spatiotemporal information, which serves as the input for subsequent processing.

[0053] Next, the fused features are fed into the multi-head attention mechanism module. The core function of this module is to perform adaptive weight calculation, that is, the model automatically learns and focuses on the important correlations between different parts of the feature sequence (such as different time points and different state parameters), thereby dynamically highlighting key information and suppressing secondary information.

[0054] Then, the output features, after being weighted by the attention mechanism and globally modeled, are fused with the original spatial features extracted from the TCN network (from its residual paths). This typically involves direct residual addition, aiming to combine the global information refined through attention with the local details preserved by the TCN, resulting in richer and more robust feature representations.

[0055] Then, the above fusion results are subjected to layer normalization to stabilize the model training process and accelerate convergence.

[0056] Ultimately, the features obtained through this series of fusion and optimization processes are the encoded features output by the encoder, which contain high-quality spatiotemporal dependencies. These features will be passed to the decoder for final trajectory prediction.

[0057] In step 105, after obtaining the encoder output, the encoder output is input into the decoder. The decoder processes the spatiotemporal perception features to obtain the flight path data of the flight equipment for a future preset time period output by the encoder.

[0058] This application embodiment captures the temporal dependence of historical track data through the GRU module, focuses on learning the spatial features of multidimensional data through the TCN-GRU module, and introduces a multi-head attention mechanism to adaptively calculate the weights of hidden features and accurately quantify and evaluate the degree of difference of various spatiotemporal features on prediction accuracy, thereby improving the accuracy of track prediction.

[0059] Based on the above embodiments, the GRU includes an update gate and a reset gate; the GRU extracts time features from historical track data to obtain time features, including: The sequence of historical flight track data is input into the GRU in time step order; For each time step, the historical track data corresponding to the current time step is combined with the hidden state of the previous time step maintained internally by the GRU to calculate the activation values ​​of the update gate and reset gate in the GRU; based on the activation values, features are extracted from the historical track data to obtain time features.

[0060] In the specific implementation process Figure 4 A schematic diagram of a GRU structure is provided for an embodiment of this application, as shown below. Figure 4 As shown, the GRU uses an update gate and a reset gate to dynamically adjust the fusion weights of historical states and current inputs to select gating information. When the track state changes significantly, the update gate can automatically reduce the weight of historical information to enhance the capture of sudden state characteristics; when the track is in a relatively stable flight state, the reset gate tends to preserve long-term historical patterns, and when encountering sudden disturbances, it lowers the reset gate value to suppress interference from irrelevant historical information. The formula is as follows:

[0061] in, This indicates an update gate that controls the historical hidden state. The proportion of information transmitted to the current state; This indicates a reset gate, determining the hidden historical state. For the current candidate state The degree of impact; among them, The multidimensional feature vector of the trajectory data at time t; Use the Sigmoid activation function; This represents the Hadamard product, an operation that allows the model to selectively fuse relevant components from historical flight track data. , , For trainable bias terms, , , These are the trainable weight values.

[0062] Flight status changes are strongly time-dependent, and these transitions are rarely instantaneous but rather exhibit a clear temporal continuity. For example, the transition from level flight to climb involves a gradual process. The GRU's gating mechanism can effectively capture the temporal patterns of these state transitions, identifying the start, duration, and end phases of the state change. Therefore, the GRU can determine the flight status of the flight equipment based on historical flight track data. Flight status can include whether there are significant changes, stable flight, or sudden disturbances. Then, the weights corresponding to the update and reset gates are determined based on the flight status. After obtaining the weights, corresponding weights are assigned to the GRU's update and reset gates, and then the temporal features are extracted from the historical flight track data.

[0063] The specific implementation method of extracting time features using GRU is as follows: First, the historical track data sequence arranged in chronological order is input into the GRU network. The GRU initializes its internal "hidden state," which serves as a memory unit and is continuously updated as time steps progress, used to carry and transmit summaries of historical information.

[0064] For each time step in the sequence (i.e., each waypoint data point), perform the following operations: The feature vector of the current time step's track point (containing information such as longitude, latitude, speed, and altitude at that moment) is combined with the "hidden state" maintained internally by GRU from the previous time step, and used as the basis for the current calculation.

[0065] Based on the combined information above, GRU calculates two gating signals (activation values) through its inherent "update gate" and "reset gate" mechanisms. Both values ​​are vectors between 0 and 1. Update gate: determines how much hidden state information from the previous time step needs to be retained in the current time step.

[0066] Reset gate: Determines how to combine the previous hidden state with the current input to generate a new candidate hidden state.

[0067] The calculated reset gate signal controls the degree of forgetting of historical information and, combined with the current input, generates a candidate hidden state. Then, using the update gate signal, like a sliding door, the hidden state from the previous time step is selectively mixed with the current candidate hidden state to obtain a new hidden state for the current time step. This new hidden state contains both filtered historical memories and new current observation information.

[0068] After completing the iterative processing for all historical time steps, the hidden state of the last time step output by the GRU, or the sequence of hidden states for all time steps, constitutes the temporal feature extracted from the input track data. This feature effectively characterizes the dynamic evolution of track parameters over time and their long-term dependencies.

[0069] This application embodiment dynamically adjusts the fusion weights of historical track data and current input through update gates and reset gates, thereby achieving the selection of gating information. Specifically, when the track state changes significantly, the update gate can automatically reduce the weight of historical track data to enhance the capture of sudden state characteristics; when the track is in a relatively stable flight state, the reset gate tends to preserve long-term historical patterns; when encountering sudden disturbances, the reset gate is reduced to suppress interference from irrelevant historical information, thereby extracting accurate time features.

[0070] Based on the above embodiments, TCN includes an extended causal convolution module and a residual module; spatial features are extracted from historical track data through TCN to obtain spatial features, including: Historical flight track data is input into the dilated causal convolution module, and intermediate features are calculated based on the dilation factor and kernel size in the dilated causal convolution module. The intermediate features are input into the residual module to obtain the spatial features output by the residual module.

[0071] In the specific implementation process Figure 5 This is a schematic diagram of the structure of an expanded causal convolution module in a TCN provided in an embodiment of this application. Figure 6 This is a schematic diagram of the residual module structure provided in an embodiment of this application, as shown below. Figure 5 and Figure 6 As shown, TCN, a convolutional neural network designed specifically for time-series data, has the core advantage of achieving in-depth mining of multi-dimensional spatial correlation features in track data through the synergistic effect of Dilated Causal Convolution (DCC) and Residual Block (RB).

[0072] The spatial characteristics of flight path information are represented by the temporal correlation of state information such as speed and altitude. TCN utilizes dilated causal convolution to form a long-range spatial correlation model. The convolution kernel only considers information from the current time and past times to ensure the temporal causal constraints of feature extraction and prevent the leakage of future information. Mathematically, this is expressed as: Where d is the dilation factor, k is the kernel size, and X is the matrix form of the input historical trajectory data. Therefore, the historical trajectory data is input into the dilated causal convolution module, and intermediate features are calculated based on the dilation factor and kernel size in the module. After obtaining the intermediate features, they are input into the residual module to obtain spatial features.

[0073] The residual module of TCN consists of two convolutional layers. Each convolutional layer includes dilated causal convolution, weight normalization, ReLU activation function, and a Dropout module. Dropout randomly discards channels of the feature map during training, forcing the network to distribute its attention to features and preventing overfitting to a single parameter. Weight normalization decomposes the convolutional kernel weights into direction vectors. With scaling scalar ,Right now This accelerates the training convergence and stabilizes the joint optimization process of multiple trajectory parameters. When the input and output dimensions of the residual path are inconsistent, a 1×1 convolution is used for linear projection to ensure the spatial compatibility of trajectory features in the deep network.

[0074] In this embodiment, spatial features are extracted using the TCN module. By leveraging the long-range dependency modeling capability of extended causal convolution, trajectory data from different times are correlated to capture the spatial relationships between multidimensional parameters, providing reliable spatial features for subsequent trajectory prediction.

[0075] Based on the above embodiments, the decoder includes a masked multi-head attention mechanism module and a fully connected layer; the decoder processes spatiotemporal perception features to obtain flight track data of a future preset duration flight device, including: The spatiotemporal awareness features are input into the mask multi-head attention mechanism module to obtain the representation sequence output by the mask multi-head attention mechanism module; The sequence is input into the fully connected layer to obtain the track data output by the fully connected layer.

[0076] In the specific implementation, to maintain the autoregressive property, the decoder uses a masked attention mechanism to ensure that the output at time t is not based on the input after time t. The calculation formula is as follows:

[0077] in, Let i be the attention score of the i-th head. This is the generated mask.

[0078] Attention score for each head Perform a softmax operation to obtain the attention weights. The weights for future moments are then reset to zero. The attention weights are calculated as follows:

[0079] Using attention weights A weighted sum operation is performed on the values ​​to obtain the output for each head. The output for each head is calculated as follows: .

[0080] Output from multiple heads The data is concatenated and subjected to a linear transformation to generate the final output of the masked multi-head attention mechanism module: .

[0081] Using a fully connected layer (linear transformation), the decoder output D is mapped to a four-dimensional vector, which is the predicted target time series:

[0082] Where D is the output of the decoder. Y This is the final prediction result, where W is the weight matrix and b is the bias term. In this way, the model can generate complete time-series prediction results, including all the important information about the trajectory.

[0083] In this embodiment of the application, the accuracy of trajectory prediction is improved because accurate spatiotemporal perception features are extracted.

[0084] This application also provides a method for training a trajectory prediction model, as detailed below: Acquire training track data within the approach control area. Historical track data includes longitude, latitude, speed, altitude, heading, and time of zero distance. The historical flight track data is constructed into multiple sample target pairs according to the time sequence; Input the samples from the target pair into the trajectory prediction model to be trained, and obtain the prediction results output by the trajectory prediction model to be trained. Based on the prediction results and objectives, the parameters of the trajectory prediction model to be trained are optimized to obtain a well-trained trajectory prediction model.

[0085] In the specific implementation process, during the data preparation phase, key trajectory data is collected, including takeoff and landing airports, aircraft type, wake turbulence, coordinates, speed, altitude, and heading. For the collected key trajectory data, some parameters are not very relevant to trajectory prediction, such as takeoff and landing airports and aircraft type; therefore, these fields can be removed. The key trajectory data after removing useless fields is then divided into training, validation, and test sets. In this embodiment, the key trajectory data in the training set is referred to as training trajectory data.

[0086] After obtaining the training trajectory data, the Min-Max algorithm can be used to normalize the training trajectory data. The purpose of normalization is to scale the data to a specific range so as to better adapt to the training of the model.

[0087] In this embodiment of the application, the normalization range is set to between -1 and 1, and the normalization formula is: Where x is the original data, and These are the minimum and maximum values ​​in the dataset, respectively. Parameters a and b are the lower and upper bounds of the normalized range, respectively.

[0088] Using min-max normalization ensures that the data remains within a certain range, avoiding problems such as gradient explosion during model training. Normalization also accelerates model convergence and improves model stability. For normalized data, inverse normalization is required to obtain the original data when calculating prediction results.

[0089] Normalizing the training trajectory data can effectively improve the training speed of the model and increase the prediction accuracy.

[0090] During the model initialization phase, the model's encoder and decoder and related parameters are set, and hyperparameters such as learning rate, batch size, and number of iterations are configured for model training.

[0091] After completing the above configuration, the training track data is constructed into multiple sample target pairs according to the time sequence, for example: This serves as the input to the trajectory prediction model to be trained. The target of the trajectory prediction model to be trained. and This forms a sample target pair. Specifically, 50 waypoints can be used as samples to predict the longitude, latitude, speed, altitude, heading, and estimated landing time of waypoints in the next 30 time units. Input the trajectory prediction model to be trained, obtain the prediction results output by the model, and optimize the parameters of the trajectory prediction model to be trained using the gradient descent method based on the prediction results and the target, so as to obtain a trained trajectory prediction model.

[0092] Optionally, when constructing sample target pairs, to adhere to the causality of time series prediction and prevent the model from peeking into future information during training, a masking operation can be applied to the future time-step data in the samples. Specifically, in the sample sequence input to the model, the track data (such as longitude, latitude, speed, altitude, heading, UTC, etc.) corresponding to future prediction time steps will be covered or replaced by special mask symbols (such as zero values), thereby ensuring that the model's prediction at each time step depends only on current and past historical information.

[0093] During training, the MSE loss function can be calculated based on the prediction results and the target. The formula for the MSE loss function is: .

[0094] Where n represents the number of samples. This represents the prediction result of the model for the i-th sample. Let represent the true target of the i-th sample. The loss value calculated using the MSE loss function is used as a standard to evaluate the prediction accuracy. The squared term amplifies the bias, thus making the model pay more attention to these samples. The lower the loss value, the more accurate the model prediction and the better the performance.

[0095] When optimizing the parameters of a trajectory prediction model to be trained, the AdamW optimizer can be used. The AdamW optimizer has the following advantages: (1) Adaptive learning rate: The AdamW optimizer can adaptively adjust the learning rate, dynamically updating the learning rate based on the gradient of each parameter. This helps to achieve better convergence at different stages of training.

[0096] (2) Weight decay: The introduction of a weight decay mechanism helps to control the size of model parameters, reduce the risk of overfitting, and improve the generalization performance of the model.

[0097] (3) Stability and convergence speed: Compared with traditional optimizers such as SGD, AdamW has better stability and convergence speed. It can achieve good results on different tasks and datasets, reducing the need for manual adjustment of the learning rate.

[0098] After the model training is completed, the trajectory prediction model can be evaluated using a test set. The accuracy and generalization can be determined by comparing the prediction results with the actual data through error metrics.

[0099] Understandably, during training, the root mean square error and / or mean absolute error are calculated based on the prediction results and the target. The convergence status of the trajectory prediction model during the training process is determined based on the root mean square error and / or mean absolute error. Whether to stop training is determined based on the convergence status.

[0100] In this embodiment, the root mean square error (RMSE) and mean absolute error (MAE) are used as prediction error metrics to evaluate the performance of the trajectory prediction model in four dimensions: longitude, latitude, altitude, and time.

[0101] Define the output of the prediction model as , i This represents the dimension of the prediction, and the actual observed value is... For each dimension, calculate the root mean square error (RMSE) and mean absolute error (MAE) for each predicted sample, then take the average of all samples. The formula for calculating the root mean square error (RMSE) is: Where N is the number of samples, and j represents the sample index. The formula for calculating the mean absolute error (MAE) is: .

[0102] These two metrics are used to measure the predictive performance of a model across different dimensions. RMSE is more sensitive to large errors because it calculates the square of the error, while MAE better reflects the actual error in the predicted values. By comparing the RMSE and MAE values ​​of a model across different dimensions, we can identify which aspects the model performs well in and which aspects need further improvement. This helps to comprehensively evaluate the model's effectiveness and provides important quantitative data support for research.

[0103] To reduce trajectory prediction error, the AdamW optimizer, a strategy combining adaptive learning rate and weight decay, is employed. AdamW's weight decay helps control parameter complexity and improves generalization. Its update rule is as follows:

[0104]

[0105]

[0106]

[0107]

[0108] in, and These represent the first-order moment estimate and the second-order moment estimate of the gradient, respectively. This is the current gradient. and These are hyperparameters that control the decay rates of the first and second moments. It's the learning rate. It is a very small constant. It is a parameter. Weight decay coefficient.

[0109] Figure 7 This application provides a schematic diagram of a trajectory prediction model loss iteration, as shown in the embodiment. Figure 7 As shown, due to data normalization before training and the use of MSE as the loss function, the initial loss value was 0.02, and the loss value was very small after the first iteration. After 500 iterations of training, the loss value decreased rapidly. After sufficient training, both the training loss and validation loss curves of the model tended to stabilize, indicating that the initial learning rate and batch size were set appropriately. The training process reached convergence at the 300th iteration, at which point the mean squared error (MSE) of the model output decreased to approximately 0.0035. The loss values ​​on both the validation and test sets remained in a low and reasonable range, confirming that the model has good generalization performance.

[0110] Although the Transformer supports parallel output, approach control area tracks are affected by dynamic factors such as real-time weather and air traffic control instructions. Predicting aircraft tracks requires real-time fusion of the latest observation data to reduce accumulated errors. Therefore, this application's embodiment utilizes TCN-dilated causal convolution to form long-range spatial correlation modeling, and the convolution kernel only considers information from the current and past times, ensuring the temporal causal constraints of feature extraction. To avoid future information leakage, the output at time 'T+1' is used as the input at time 'T+2', which better reflects the physical continuity of aircraft motion in the approach control area. Furthermore, parameters can be shared across steps, capturing sequence changes point by point. Performance is evaluated using the first 500 points from 15227 test flights as ground truth and 500 consecutive predicted points. Figure 8 The comparison chart provided in this application embodiment shows the difference between the true value and the predicted value for a certain feature. The comparison results show that the trajectory prediction model in this application embodiment has high prediction accuracy.

[0111] To demonstrate the performance of the prediction model, this embodiment randomly selects a track from the test set for verification, showcasing only three dimensions: altitude, longitude, and latitude. The illustration shows the prediction capability of the track prediction model provided by this embodiment under different step sizes. Figure 9 , Figure 10 Through the analysis of Figure 9 and Figure 10A comparison of the three-dimensional results of the trajectory prediction model under different step lengths shows that the goodness of fit of the model increases with the increase of the prediction step length. This phenomenon fully confirms the significant advantage of the model in long-term prediction scenarios, that is, the trajectory prediction model can accurately capture and effectively predict the evolution trend of aircraft trajectories in multiple spatial dimensions such as longitude, latitude, and altitude.

[0112] To verify the feasibility of the proposed spatiotemporal recognition Transformer (TA-TFT) model in this application, it was compared with traditional models GRU, Bi-LSTM, Transformer (C-TFT), and more advanced trajectory prediction models HNNM and FT-TF. The training set, validation set, test set, and error evaluation criteria were kept identical. The optimal parameter values ​​of each model were compared with these models. To comprehensively evaluate the prediction results, RMSE and MAE were selected as error indices to evaluate the four dimensions of the sample features. RMSE and MAE were calculated separately as observation dimension error indices to reflect the measurement results of the longitude, latitude, altitude, and time dimensions, facilitating the analysis of performance in each dimension and model adjustment.

[0113] Figure 11 and Figure 12 The paper demonstrates the prediction errors of each model at different time steps. To meet the ICAO requirements for approach control area track update intervals, the track prediction model predicts tracks for the next 250 seconds with a step size of 50, conforming to the ICAO requirement of a prediction time window of ≥120 seconds. Prediction accuracy evaluation shows that the basic GRU model has the lowest accuracy. The Long Short-Term Memory Network (Bi-LSTM) effectively captures the long-term dependencies of the sequence, and its prediction accuracy is higher than that of GRU. The HNNM and FT-TF models, which perform spatiotemporal feature learning, also show relatively ideal accuracy. However, for spatiotemporal recognition, the track prediction model (TA-TFT) in this embodiment of the application has significantly higher prediction accuracy than the basic GRU, Bi-LSTM, traditional Transformer (C-TFT), and HNNM and FT-TF models.

[0114] The above experimental results further verify that the model has good performance, for the following reasons: 1. Multi-feature set: Utilize a set of features that contain more valuable aviation features, such as departure airport, flight type, latitude and longitude, altitude, time, etc., to improve the model's recognition of flight trajectory features.

[0115] 2. In the encoder, GRU is a recurrent neural network that processes time series. The GRU structure simplifies the gating mechanism. Compared with LSTM, it can model long-term dependencies in a similar way, while further increasing the computational efficiency and solving the gradient vanishing problem in long sequences very well.

[0116] 3. In the embodiments of this application, the TCN module, based on the spatial characteristics of approach control area tracks, uses the long-range dependency modeling capability of extended causal convolution to associate longitude, latitude and altitude changes at different times, and captures the spatial correlation between multidimensional parameters.

[0117] 4. Multi-head self-attention mechanism: The self-attention mechanism in the Transformer model addresses long-term temporal dependencies between sequences and interactions between features. It assigns weights based on the similarity between the query and the key, which helps to indicate the fitting effect between different time steps and different features.

[0118] In summary, the trajectory prediction method proposed in this application can learn the flight characteristics of different airports simply by importing data and training it independently. Furthermore, since the peak and off-peak hours of airports are mostly fixed, providing sufficient datasets during training is sufficient for trajectory prediction in approach control areas, thus providing reliable technical support for long-term trajectory monitoring in approach control areas.

[0119] This application also provides ablation experiments. The trajectory prediction model removed the GRU module, TCN module, and multi-head attention mechanism module that extracts basic trajectory data. All ablation experiments were conducted with a step size of 50, and the RMSE and MAE of each model were calculated. The experimental results are shown in the table below.

[0120]

[0121] Ablation experiments show that the complete TA-TFT model (i.e., the trajectory prediction model in this embodiment) performs best in predictions across all dimensions of longitude, latitude, altitude, and time. Its average RMSE is 31.7% lower than the incomplete model, and its average MAE is also 28.6% lower. This indicates that GRU, TCN, and the attention mechanism are crucial for improving model performance; their synergistic effect enhances the model's ability to extract and predict trajectory features, significantly improving prediction accuracy.

[0122] Figure 13 This is a schematic diagram of a trajectory prediction device provided in an embodiment of this application. The device can be a module, program segment, or code on an electronic device. It should be understood that this device is similar to the one described above. Figure 1 The method implementation corresponds to this and can be executed. Figure 1The specific functions of the device involved in the various steps of the method embodiment can be found in the description above; to avoid repetition, detailed descriptions are omitted here. The device includes: a data acquisition module 1301, a temporal feature extraction module 1302, a spatial feature extraction module 1303, a feature combination module 1304, and a prediction module 1305, wherein: The data acquisition module 1301 is used to acquire historical flight track data of the flight equipment within the approach control area during a preset historical time period; The time feature extraction module 1302 is used to extract time features from the historical track data using GRU to obtain time features; the time features are used to characterize the dynamic evolution of the historical track data over time. The spatial feature extraction module 1303 is used to extract spatial features from the historical flight track data through TCN to obtain spatial features; the spatial features are used to characterize the temporal correlation of the spatial state of the flight equipment. Feature integration module 1304 is used to process the temporal features and spatial features of the multi-head attention mechanism module to obtain spatiotemporal perception features; The prediction module 1305 is used to process the spatiotemporal perception features through the decoder to obtain the flight track data of the flight equipment for a future preset duration.

[0123] Based on the above embodiments, the GRU includes an update gate and a reset gate; the time feature extraction module 1302 is specifically used for: The sequence of historical flight track data is input into the GRU in time step order; For each time step, the historical track data corresponding to the current time step is combined with the hidden state of the previous time step maintained internally by the GRU, and the activation values ​​of the update gate and reset gate in the GRU are calculated. Based on the activation value, feature extraction is performed on the historical flight track data to obtain the time feature.

[0124] Based on the above embodiments, the TCN includes an extended causal convolution module and a residual module; the spatial feature extraction module 1303 is specifically used for: The historical flight track data is input into the dilated causal convolution module, and intermediate features are calculated based on the dilation factor and kernel size in the dilated causal convolution module. The intermediate features are input into the residual module to obtain the spatial features output by the residual module.

[0125] Based on the above embodiments, the decoder includes a masked multi-head attention mechanism module and a fully connected layer; the feature joint module 1304 is specifically used for: The spatiotemporal awareness features are input into the mask multi-head attention mechanism module to obtain the representation sequence output by the mask multi-head attention mechanism module; The representation sequence is input into the fully connected layer to obtain the track data output by the fully connected layer.

[0126] Based on the above embodiments, the device further includes a model training module, used for: Acquire training track data in the approach control area, the training track data including longitude, latitude, speed, altitude, heading and time at zero distance; The training trajectory data is constructed into multiple sample target pairs according to the time sequence; The samples in the target pair are input into the trajectory prediction model to be trained, and the prediction results output by the trajectory prediction model to be trained are obtained. Based on the prediction results and the target, the parameters of the trajectory prediction model to be trained are optimized to obtain the trained trajectory prediction model.

[0127] Based on the above embodiments, the device is also used for: During training, the root mean square error and / or mean absolute error are calculated based on the prediction results and the target. The convergence state of the trajectory prediction model during the training process is determined based on the root mean square error and / or the mean absolute error. Whether to stop training is determined based on the convergence status.

[0128] Based on the above embodiments, the device further includes a normalization module, used for: The training trajectory data was normalized using the Min-Max algorithm.

[0129] Figure 14 This is a schematic diagram of the physical structure of the electronic device provided in the embodiments of this application, such as... Figure 14 As shown, the electronic device includes: a processor 1401, a memory 1402, and a bus 1403; wherein: The processor 1401 and the memory 1402 communicate with each other through the bus 1403; The processor 1401 is used to call program instructions in the memory 1402 to execute the methods provided in the above-described method embodiments, including, for example,: acquiring historical flight track data of the flight equipment within an approach control area during a preset historical time period; extracting time features from the historical flight track data using a GRU to obtain time features; the time features are used to characterize the dynamic evolution of the historical flight track data over time; extracting spatial features from the historical flight track data using a TCN to obtain spatial features; the spatial features are used to characterize the temporal correlation of the spatial state of the flight equipment; processing the time features and spatial features through the multi-head attention mechanism module to obtain spatiotemporal perception features; and processing the spatiotemporal perception features through the decoder to obtain flight track data of the flight equipment for a future preset duration.

[0130] Processor 1401 can be an integrated circuit chip with signal processing capabilities. The processor 1401 can be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it can also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. It can implement or execute the various methods, steps, and logic block diagrams disclosed in the embodiments of this application. The general-purpose processor can be a microprocessor or any conventional processor.

[0131] The memory 1402 may include, but is not limited to, random access memory (RAM), read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), etc.

[0132] This embodiment discloses 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 the program instructions are executed by the computer, the computer can execute the methods provided in the above-described method embodiments, such as: acquiring historical track data of a flight device within a preset historical time period in an approach control area; extracting time features from the historical track data using a GRU to obtain time features; the time features are used to characterize the dynamic evolution of the historical track data over time; extracting spatial features from the historical track data using a TCN to obtain spatial features; the spatial features are used to characterize the temporal correlation of the spatial state of the flight device; processing the time features and spatial features through the multi-head attention mechanism module to obtain spatiotemporal perception features; and processing the spatiotemporal perception features through the decoder to obtain track data of the flight device for a future preset duration.

[0133] This embodiment provides a non-transitory computer-readable storage medium storing computer instructions that cause the computer to execute the methods provided in the above-described method embodiments. These instructions include, for example: acquiring historical flight track data of a flight device within an approach control zone over a preset historical time period; extracting temporal features from the historical flight track data using a GRU to obtain temporal features; the temporal features characterizing the dynamic evolution of the historical flight track data over time; extracting spatial features from the historical flight track data using a TCN to obtain spatial features; the spatial features characterizing the temporal correlation of the flight device's spatial state; processing the temporal and spatial features using a multi-head attention mechanism module to obtain spatiotemporal awareness features; and processing the spatiotemporal awareness features using a decoder to obtain flight track data of the flight device for a future preset duration.

[0134] In the embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. The apparatus embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. Furthermore, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Additionally, the displayed or discussed mutual couplings, direct couplings, or communication connections may be through some communication interfaces; indirect couplings or communication connections between devices or units may be electrical, mechanical, or other forms.

[0135] Furthermore, the units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0136] Furthermore, the functional modules in the various embodiments of this application can be integrated together to form an independent part, or each module can exist independently, or two or more modules can be integrated to form an independent part.

[0137] In this document, relational terms such as first and second are used only to distinguish one entity or operation from another entity or operation, without necessarily requiring or implying any such actual relationship or order between these entities or operations.

[0138] The above description is merely an embodiment of this application and is not intended to limit the scope of protection of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A method of track prediction, characterized by, The method is applied to a track prediction model, and the track prediction model is a Transformer architecture model; the track prediction model comprises an encoder and a decoder; the encoder comprises an integrated gated recurrent unit (GRU), a temporal convolutional network (TCN), and a multi-head attention mechanism module; and the method comprises the following steps: acquiring historical track data of a flight device in an approach control area within a preset historical time period; extracting time features of the historical track data by using the GRU, so as to obtain the time features; the time features are used to represent a dynamic evolution law of the historical track data over time; extracting spatial features of the historical track data by using the TCN, so as to obtain the spatial features; the spatial features are used to represent time correlation of a spatial state of the flight device; processing the time features and the spatial features by using the multi-head attention mechanism module, so as to obtain spatio-temporal perception features; processing the spatio-temporal perception features by using the decoder, so as to obtain track data of the flight device in a future preset time period.

2. The method of claim 1, wherein, The GRU comprises an update gate and a reset gate; the step of extracting the time features of the historical track data by using the GRU comprises the following steps: inputting a sequence of the historical track data into the GRU in a time step order; for each time step, combining historical track data corresponding to the current time step with a hidden state of a previous time step maintained in the GRU, to calculate activation values of the update gate and the reset gate in the GRU; extracting features of the historical track data based on the activation values, to obtain the time features.

3. The method of claim 1, wherein, The TCN comprises a dilated causal convolution module and a residual module; the step of extracting the spatial features of the historical track data by using the TCN comprises the following steps: inputting the historical track data into the dilated causal convolution module, to calculate and obtain intermediate features based on a dilation factor and a convolution kernel size in the dilated causal convolution module; inputting the intermediate features into the residual module, to obtain the spatial features output by the residual module.

4. The method of claim 1, wherein, The decoder comprises a masked multi-head attention mechanism module and a fully connected layer; the step of processing the spatio-temporal perception features by using the decoder to obtain the track data of the flight device in the future preset time period comprises the following steps: inputting the spatio-temporal perception features into the masked multi-head attention mechanism module, to obtain a representation sequence output by the masked multi-head attention mechanism module; inputting the representation sequence into the fully connected layer, to obtain the track data output by the fully connected layer.

5. The method according to any one of claims 1 to 4, characterized in that, The method further comprises the following steps: acquiring training track data in an approach control area, wherein the training track data comprises longitude, latitude, speed, height, heading, and time of distance zero; constructing the training track data into a plurality of sample target pairs according to a time sequence; inputting samples in the sample target pairs into a track prediction model to be trained, to obtain a prediction result output by the track prediction model to be trained; optimizing parameters of the track prediction model to be trained according to the prediction result and the target pairs, to obtain a trained track prediction model.

6. The method of claim 5, wherein, The method further comprises the following steps: calculating a root mean square error and / or a mean absolute error according to the prediction result and the target during the training process; judging a convergence state of the track prediction model in the training process according to the root mean square error and / or the mean absolute error; determining whether to stop the training according to the convergence state.

7. The method of claim 5, wherein, After obtaining the training track data in the approach control area, the method further comprises: normalizing the training track data by using a Min-Max algorithm.

8. An electronic device, comprising: comprise: a processor, a memory and a bus, wherein: the processor and the memory complete communication with each other through the bus; the memory stores program instructions capable of being executed by the processor, and the processor calling the program instructions can execute the method according to any one of claims 1-7.

9. A non-transitory computer-readable storage medium, comprising: The non-transitory computer readable storage medium stores computer instructions, and the computer instructions make the computer execute the method according to any one of claims 1-7 when the computer runs.

10. A computer program product, characterised in that, comprise computer program instructions, and the computer program instructions are read and run by the processor and execute the method according to any one of claims 1-7.

Images