4D track prediction method based on prior track guidance and multivariable time lag dynamic adjustment

Abstract

The invention discloses a 4D track prediction method based on prior track guidance and multivariable time lag dynamic adjustment, which comprises the following steps: firstly, preprocessing real historical flight data and meteorological data; extracting prior track data of a fixed navigation point from a flight navigation database, and generating prior features; meanwhile, internal source features are extracted from internal source variables of the preprocessed real historical flight data; extracting exogenous features from the exogenous variables of the preprocessed real historical flight data; inputting the priori features and the endogenous features into a priori feature guide cross attention module to generate priori fused endogenous features; inputting the endogenous feature and the exogenous feature of the fusion prior into an exogenous cross attention module to generate a final fusion feature; and the final fusion feature is spliced with the historical track sequence, and a future 4D track prediction result is further generated through linear mapping. According to the method, the accuracy of track prediction is improved through a method of automatically aligning time lags possibly existing between multiple variables.

Description

Technical Field

[0001] This invention relates to the field of trajectory prediction in time series analysis, specifically a 4D trajectory prediction method based on prior trajectory guidance and multivariate time lag dynamic adjustment. Background Technology

[0002] Against the backdrop of continued growth in global air transport demand, traditional air traffic management models are facing increasingly severe challenges. With the continuous increase in flight density, problems such as airspace resource shortages, operational efficiency bottlenecks, and safety risks under complex weather conditions are becoming more prominent, giving rise to the need for accurate prediction of aircraft four-dimensional (4D) tracks (covering four dimensions: longitude, latitude, altitude, and time).

[0003] The primary value of 4D flight path prediction lies in enhancing the precision of airspace operations. Traditional management methods, relying on fixed routes and rules, struggle to cope with real-time changes in traffic flow and complex flight environments. 4D prediction, however, provides air traffic management systems with precise flight dynamic information, enabling controllers and airlines to plan flight paths in advance and optimize takeoff and landing sequences. This effectively alleviates congestion in busy airspaces and reduces ground taxiing and in-flight waiting times.

[0004] In recent years, time series forecasting technology has developed rapidly under the drive of deep learning. From traditional statistical models to machine learning models and now to deep learning models represented by Transformer, it has broken through the bottlenecks of long sequence dependency capture, multimodal data fusion and cross-domain generalization ability, and has significantly improved in processing high-dimensional time series data, dynamic environment adaptability and prediction accuracy.

[0005] Traditional statistical models (such as ARIMA and SARIMA) have significant shortcomings in trajectory prediction: they rely on linear assumptions and fixed model structures; they have weak adaptability to data non-stationarity; their strict prior assumptions are incompatible with outliers, missing values, and multi-source heterogeneity characteristics commonly found in trajectory data; they rely on manual feature engineering to integrate information, and when generalizing across scenarios, parameters need to be manually adjusted for different airports, aircraft types, etc., resulting in high migration costs and insufficient flexibility, making it difficult to meet the requirements of modern trajectory prediction for high accuracy, real-time performance, and robustness.

[0006] Machine learning models (such as Random Forest and SVM) have several limitations in trajectory prediction: they rely on manually designed feature engineering, making it difficult to automatically capture deep correlations in high-dimensional spatiotemporal data; their ability to model complex nonlinear dynamics is limited by fixed model structures, requiring empirical parameter tuning for approximate fitting, resulting in insufficient flexibility; in temporal dependency analysis, they rely on preset time windows to process sequence information, and cannot automatically learn long-distance dependencies like deep learning models; in terms of interpretability, the decision logic of complex models is obscure, making it difficult to meet the causal attribution requirements that are crucial for safety in trajectory prediction, while the expressive power of lightweight models is insufficient to cope with high-dimensional nonlinear scenarios.

[0007] In recent years, an increasing number of deep learning models have been applied to time series forecasting. Deep learning models, with their ability to automatically model high-dimensional features, have become the mainstream models in various fields. Convolutional Neural Networks (CNNs) were first applied to image processing. By adjusting the size of the convolutional kernels, the receptive field of the model to image data can be changed, allowing the model to learn how to extract global and local features. However, in time series forecasting, CNNs rely on local convolutional kernels, making it difficult to effectively capture long-distance temporal dependencies. Their translation invariance leads to insufficient sensitivity to sequence order, and they lack mechanisms for explicitly modeling multivariate cross-dimensional interactions, exhibiting poor adaptability to pattern changes in non-stationary data. Recurrent Neural Networks (RNNs) and their typical variants, Long Short-Term Memory (LSTM) and Gated Logic Regulators (GRUs), are among the earliest deep learning sequence models. Original RNNs are prone to the vanishing gradient problem. Later models like LSTM and GRU introduced gating mechanisms, which only improved this problem to some extent; the vanishing gradient problem still exists for increasingly longer datasets. Furthermore, due to the recursive modeling characteristics of their networks, RNNs also suffer from low training efficiency.

[0008] Introduced by Google in 2017, the Transformer model has been widely studied by scholars due to its excellent modeling capabilities for long series. Transformer-like models have been continuously proposed in the field of time series forecasting, such as Informer, Autoformer, FEDformer, and Itransformer. Informer utilizes a sparse attention mechanism, while Autoformer uses an autocorrelation-based attention mechanism to reduce the time complexity of attention computation. FEDformer's optimization relies more on frequency domain transformation and random sampling, further reducing the complexity to O(N), and its global frequency domain modeling enhances its ability to capture long-term trends. Itransformer, through a transposed Transformer module, models the correlation between multiple variables, better utilizing external information introduced by other variables for time series forecasting. However, for long-term 4D trajectory forecasting, these models do not effectively utilize valuable prior trajectory information, and they also lack a good solution for the potential time lag of multiple variables. Summary of the Invention

[0009] The purpose of this invention is to overcome the shortcomings of existing technologies and propose a long-term 4D trajectory prediction method based on prior trajectory guidance and combined with multivariate time lag dynamic adjustment. Recognizing that prior trajectories in flight navigation databases are not well utilized and that there may be time lags between meteorological data such as temperature, humidity, and wind speed, this invention employs a method that uses trajectory data from a navigation database conforming to the ARINC 424 standard as prior knowledge and automatically aligns the possible time lags between multiple variables using time lag convolution to improve the accuracy of trajectory prediction.

[0010] To achieve the above objectives, the technical solution specifically adopted by the present invention is as follows:

[0011] A 4D trajectory prediction method based on prior trajectory guidance and multivariate time lag dynamic adjustment includes the following steps:

[0012] Step 1: Preprocess the real historical flight data and meteorological data;

[0013] Step 2: Extract prior track data of fixed navigation points from the flight navigation database and input it into the prior feature extraction module to generate prior features; at the same time, input the internal variables of the preprocessed real historical flight data into the internal feature extraction module to obtain internal features;

[0014] Step 3: Input the exogenous variables of the preprocessed real historical flight data into the adaptive time-delay convolution exogenous feature encoding module to generate exogenous features;

[0015] Step 4: Input the prior features and intrinsic features into the prior feature-guided cross-attention module, and perform temporal fusion with the prior features as a guide to generate the fused prior intrinsic features;

[0016] Step 5: Input the intrinsic and extrinsic features of the fusion prior into the extrinsic cross-attention module to perform variable-level feature fusion and generate the final fused features;

[0017] Step 6: Concatenate the final fused features with the historical trajectory sequence after the endogenous feature extraction module, and generate the future 4D trajectory prediction result in one step through linear mapping.

[0018] Preferably, the preprocessing method includes: outlier handling, trackpoint interpolation, temporal feature standardization, and data alignment.

[0019] Preferably, in step 2, the operation of the prior feature extraction module includes:

[0020] Perform a patching operation on the prior track data to divide it into segments of length P;

[0021] The fragment is linearly mapped to a high-dimensional feature and concatenated with a learnable global token.

[0022] Global trajectory features are extracted using a self-attention mechanism, and prior features are output.

[0023] Preferably, the endogenous variables include longitude, latitude, altitude, and time.

[0024] Preferably, the external variables include flight number, temperature, humidity, and wind speed.

[0025] Preferably, in step 2, the operation of the intrinsic feature extraction module is the same as that of the prior feature extraction module. The input of the intrinsic feature extraction module is to stitch multiple real historical tracks into a sequence by splicing.

[0026] Preferably, in step 2, the operation of the adaptive time-delay convolutional exogenous feature encoding module includes:

[0027] First convolutional layer: First, adjust the dimensionality of the n historical track data with k external variables to [value]. Intermediate results were obtained using the grouped Conv1d method, and their dimensions were adjusted to obtain periodic features.

[0028] Perform a Gumbel-Softmax operation on the obtained periodic feature C, and set hard=True to convert the output into a strict one-hot vector;

[0029] The second convolutional layer uses a time-lag window kernel that employs a strictly one-hot vector as the kernel for the second convolutional layer.

[0030] The second convolutional layer first processes X... ex Padding is performed only at the end of the L-axis, using a moving average method. The padding length is l-1, adjusted for dimension variation.

[0031] by Grouped convolution is used for the convolution kernel, for X padex After performing the convolution operation, adjust the dimensions to obtain That is, to perform time-lag alignment on k external variables of n historical tracks;

[0032] Finally The last two dimensions are flattened into one dimension, and a linear layer is used to map the flattened dimension size from n*L to the hidden representation dimension size d_model to obtain the final external feature.

[0033] Preferably, in step 4, the operation of the prior feature-guided cross-attention module includes:

[0034] Based on prior features As a query, internal features As Key and Value;

[0035] By calculating similarity through cross-attention, the physical constraints of prior tracks are incorporated into historical track features to obtain intrinsic features that fuse prior information.

[0036] Preferably, in step 5, the operation of the exogenous cross-attention module includes:

[0037] Integrate prior intrinsic features As a query, external features As the key and value, they undergo another cross-attention mechanism, followed by layer normalization, a feedforward layer, and another layer normalization to obtain the final fused feature.

[0038] Preferably, step 6 employs a one-step generative prediction, specifically including:

[0039] Final output feature O finall Compared with historical tracks after endogenous feature extraction module By splicing After flattening the last two dimensions, a linear layer mapping operation is used to further obtain the dimension as follows. The tensor is the predicted trajectory sequence.

[0040] Preferably, the endogenous feature extraction module additionally incorporates temporal encoding:

[0041] The time characteristics of historical flight tracks are broken down into year, month, day, hour, minute, and second components;

[0042] The location information is supplemented by encoding through the Embedding layer and then superimposed onto the track features.

[0043] This invention has the following characteristics and beneficial effects:

[0044] The above technical solution, when incorporating prior trajectory data, offers several advantages for trajectory prediction using historical trajectory data: First, prior trajectory data compensates for potential missing physical constraints in real historical trajectory data, enabling the model to learn spatiotemporal patterns that better align with actual flight logic and avoiding unreasonable predictions that violate flight principles. Second, the implicit trajectory priors in the prior trajectory data enhance the consistency of predicted trajectories, ensuring the continuity and rationality of predicted trajectories across different flight phases. Finally, the combination of prior and historical data forms a hybrid modeling framework of "data-driven + knowledge-guided," improving the model's generalization ability in scenarios with few samples, while providing interpretable physical evidence for the prediction results, meeting the stringent safety and reliability requirements of the aviation field.

[0045] Time lag alignment of multiple auxiliary predictive variables can significantly improve the accuracy and reliability of flight path prediction. Its core benefit stems from the accurate modeling of the causal temporal relationships between variables: First, by identifying and aligning the lag effects between different variables such as meteorological data, airspace traffic, and air traffic control instructions, spurious correlations or omissions of key information caused by time misalignment can be avoided. This allows the model to capture the true physical transmission mechanisms, such as how changes in air pressure affect flight attitude adjustments after a delay in the response of aircraft sensors. Time lag alignment reduces redundancy and noise in the feature space. By eliminating meaningless leading / lagging combinations, such as incorrectly using future weather data for current flight path prediction, the model is forced to focus on valid inputs that conform to temporal causality, such as the impact of runway queue length at time t-1 on takeoff time at time t. Especially when dealing with multi-source heterogeneous data, such as real-time radar signals and flight plans several hours in advance, timestamp calibration can be used to construct logically consistent training samples, improving the model's learning efficiency for complex spatiotemporal dependencies. Furthermore, aligned lag characteristics enhance the interpretability of prediction results. For example, cross-correlation analysis can determine the optimal lag order for wind speed and flight path deviation, providing controllers with a clear path of variable influence and aiding in the design and validation of decision support systems. Ultimately, this refined time alignment mechanism enables models to more accurately simulate the temporal transmission effects between variables when dealing with dynamic aviation environments such as cascading delays caused by sudden weather events. This results in predictions that better reflect actual flight logic, improving the collaborative decision-making capabilities and operational safety of air traffic control systems. Attached Figure Description

[0046] Figure 1 This is a flowchart illustrating the overall workflow of this invention.

[0047] Figure 2 This is a schematic diagram of the prior feature extraction module in this invention.

[0048] Figure 3 This is a schematic diagram of the feature extraction module for intrinsic variables of historical flight tracks in this invention.

[0049] Figure 4 This is a schematic diagram of the adaptive time-delay convolutional external feature encoding module in this invention.

[0050] Figure 5 This is a schematic diagram of the attention module that guides crossover based on prior features in this invention. Detailed Implementation

[0051] The present invention will now be described in detail with reference to specific embodiments. These embodiments will help those skilled in the art to further understand the present invention, but do not limit the invention in any way. It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other.

[0052] A 4D trajectory prediction method based on prior trajectory guidance and multivariate time lag dynamic adjustment is proposed. The aim is to address the shortcomings of existing methods in the field of time series prediction, focusing on the lack of standard trajectory guidance and the potential time lag between exogenous variables in real-world scenarios. The method involves extracting features from prior trajectory points and integrating them into historical real trajectory features. Then, the features extracted from exogenous variables, processed by a convolutional time lag alignment method, are again integrated into the historical trajectory features, and finally, future trajectory prediction is performed.

[0053] For feature extraction and fusion, this method adopts the self-attention mechanism using learnable global tokens and the cross-attention mechanism of internal and external variables proposed by the TimeXer model. The overall model consists of a prior feature extraction module, a historical track internal variable feature extraction module, an adaptive time-delay convolution external feature encoding module, and a prior feature-guided cross-attention module. It is particularly emphasized that this method focuses on using prior track points to guide the alignment with the time lag of external variables, rather than feature fusion.

[0054] like Figure 1 As shown, the specific steps are as follows:

[0055] Step 1: Preprocess the real historical flight data and meteorological data.

[0056] Specifically, considering the longitude, latitude, altitude, and time of historical flight tracks as endogenous variables, it can be expressed as follows: This represents n tracks, each track sequence having a length of L and containing 4 track variables. This represents one of the historical shipping routes. This represents a time step within a flight path, and a time step contains four variables; variables such as flight number, speed, heading, rate of climb or descent, temperature, humidity, and wind speed are considered as exogenous variables. This represents n tracks, each track sequence having length L and containing track variables k.

[0057] The required flight longitude, latitude, and altitude waypoints will be extracted from the ARINC 424 navigation database as prior waypoint data. This represents one track, with a track sequence length of L and containing 3 track variables. To address the issue of varying time intervals between historical track points, a cubic spline interpolation method is employed to reconstruct track features. The same sampling frequency is set for each track, ensuring equal time intervals between track points. Track points within the same time range as historical track points are extracted from prior track data and sampled at the same frequency to obtain the same number of time steps as historical track points. This avoids interference to the model caused by track points within invisible time ranges introduced from historical track data.

[0058] To eliminate unit differences, smooth gradients between different batches and layers of data, and accelerate model convergence, a standardization (z-score) method was used to process longitude, latitude, altitude, velocity, and meteorological data. This method scales the data to a standard normal distribution, thereby eliminating dimensional differences and ensuring that the weights of different features are calculated and compared fairly.

[0059] Step 2: Extract prior track data of fixed navigation points from the flight navigation database and input it into the prior feature extraction module to generate prior features; at the same time, input the internal variables of the preprocessed real historical flight data into the internal feature extraction module to obtain internal features.

[0060] In this embodiment, the flight navigation database is a navigation database that conforms to the ARINC 424 standard.

[0061] Furthermore, the method for extracting prior features is as follows:

[0062] The prior waypoints in the navigation database conforming to the ARINC 424 standard are input into the prior feature extraction module. After the patching operation, the patched sequence is obtained and concatenated with the learnable global token in the patch dimension. Then, self-attention calculation is performed. Through training, the features of the prior waypoints are automatically learned into the global token.

[0063] The specific implementation process is as follows:

[0064] The prior feature extraction module in this invention, such as Figure 2 As shown, in time series prediction scenarios, accurately capturing the temporal dependency features of track variables is a core prerequisite for constructing a self-attention mechanism. A patch-based representation method is adopted, with each patch mapped to a high-dimensional representation. Furthermore, a learnable global token is introduced for each endogenous variable, serving as a macroscopic representation for interaction with prior and historical tracks. Let L represent the sequence length of the input track, and P represent the patch length. This indicates the number of patches in the partition.

[0065] First, let the input dimension be... The prior track points are used to obtain a tensor on the sequence dimension L through patching operations. After linear mapping, we get Among them, K priori It is an intermediate result obtained by patching the prior track and then mapping the third dimension P to the size d_model.

[0066] The next step and a dimension A learnable token can be concatted to obtain a tensor. Then, self-attention, residual connections, and layer normalization are performed. These operations extract the longitude, latitude, and altitude features from a prior track into prior features. The middle way guides us to the true course of history.

[0067] The specific prior feature extraction process can be expressed by the formula:

[0068]

[0069] K priori =Project(V priori )

[0070] SA priori =Self-Attention([K priori G priori ])

[0071]

[0072] It should be noted that the parameters in the above formulas are the results obtained from prior track patching, the result of mapping the last dimension P to the size d_model, and the result of self-attention. All of these can be considered intermediate results; the prior feature extraction module ultimately needs the prior features.

[0073] Furthermore, the method for extracting endogenous features is as follows:

[0074] The extraction method of intrinsic features is similar to that of prior features. The difference is that prior track data contains only one track, while historical tracks contain multiple tracks and are spliced ​​into a sequence. The spliced ​​historical track point data is then input into the intrinsic feature extraction module to obtain intrinsic features.

[0075] The specific implementation process is as follows:

[0076] The endogenous feature extraction module in this invention, such as Figure 2As shown, the dimensions of the input historical flight track data are adjusted to... Here, n represents n real historical tracks. Merging these n historical tracks results in an n*L sequence dimension. Subsequent operations are the same as the prior feature extraction method, but with the addition of processing for temporal features, thus incorporating the temporal features of the historical track data. The data is split to obtain multiple time components including year, month, day, week, hour, minute, and second. After embedding, these components are added to the source variables of the historical track as a supplement to the position encoding. Finally, it is combined with a dimension... The learnable global token is obtained by performing a concat operation. Through self-attention operations, the system learns the longitude, latitude, altitude, and time characteristics of historical real-world flight paths, thereby extracting intrinsic features. Furthermore, the learnable global token information is incorporated into the historical track K through LayerNorm normalization. en In the process, the historical flight path processed by the intrinsic feature extraction module is obtained.

[0077] Specifically, the process of extracting intrinsic features can be expressed by the following formula:

[0078] V en =Patching(X) en )

[0079] K en =Project(V en )+Embed(X time )

[0080] SA en =Self-Attention([K en G en ])

[0081]

[0082] Step 3: Input the exogenous variables of the preprocessed real historical flight data into the adaptive time-delay convolution exogenous feature encoding module to generate exogenous features.

[0083] In this embodiment, historical track data and its external variables are input into an adaptive time-delay convolutional external feature encoding module. The input data passes through two convolutional layers. The first convolutional layer performs convolution operations on the external features, extracting periodic features from the input data using the receptive field of the convolutional kernel. The multiple output channels are then processed using Gumbel-Softmax to generate sparse shifted convolutional kernels with the same number of output channels, such as [1,0,0,0,0,...,0] and [0,1,0,0,0,...,0]. The second convolutional layer uses the generated sparse shifted convolutional kernels to perform time-delay convolution on the historical track data. Then, a linear layer is used to select from the multiple outputs for each external variable to align any potential time lag in the external variables. Finally, a learnable global token is used to extract features for each aligned external variable to obtain the final external features.

[0084] Adaptive time-delay convolutional exogenous feature encoding module, such as Figure 4 As shown, it is understandable that the main purpose of exogenous variables is to provide information beyond the track data itself for track prediction. In track prediction, the longitude, latitude, and altitude of track points are considered endogenous variables, while the remaining variables are exogenous variables. When extracting periodic information, convolution, with its local receptive field and weight sharing mechanism, can directly capture fixed periodic patterns by adapting the kernel size to the period length. Its translation invariance ensures that the same periodic pattern is consistently extracted at different locations.

[0085] Specifically, the dimensions of the n historical track data with k external variables are first adjusted to... Periodic features are obtained using a grouped Conv1d method, with k*n groups. The Conv1d layer has k*n input channels and k*n output channels. The kernel size is determined according to the formula... The adjustment is performed, where l is the length of the second convolutional layer kernel and also the size of the adjustment window for time lag. After passing through the first convolutional layer, an intermediate result is obtained, and its dimensions are adjusted to obtain the periodic features of the exogenous variable. The obtained periodic feature C is subjected to a Gumbel-Softmax operation, and setting hard=True converts the output into a strict one-hot vector, which is used as the time-lag window convolution kernel for the second convolution layer. The final dimension takes the form [[0,0,1,…,0]]. The Gumbel-Softmax method transforms the discrete sampling process into a differentiable operation through continuous relaxation, enabling the neural network to learn the position selection of discrete structures such as convolutional kernels end-to-end.

[0086] The second convolutional layer first processes X... exPadding is performed only at the end of the L-axis, using a moving average method. The padding length is l-1, adjusted for dimension variation. With time lag window convolution kernel C kernel The convolution kernel uses grouped convolution, with k*n groups, k*n input channels, and k*n output channels. For X... padex After performing the convolution operation, adjust the dimensions to obtain That is, the result of time-lag alignment of k external variables for n historical tracks.

[0087] It should be noted that, It is the result of the adaptive time-delay convolution of the external feature modules, that is, the result of time-delay alignment of k external variables from n historical tracks. Only one more Flatten and mapping operation is needed to obtain the final external features. It can be considered an intermediate result.

[0088] Finally The last two dimensions are flattened into one dimension, and a linear layer is used to map the flattened dimension size from n*L to the hidden representation dimension size d_model to obtain the final external feature.

[0089] C kernel =Gumbel-Softmax(Conv1d(X) ex ))

[0090]

[0091] Understandably, the last two dimensions refer to the last two dimensions of this feature, namely n and L dimensions.

[0092] Specifically, the original data is 3-dimensional, with the first dimension of size k, the second dimension of size n, and the third dimension of size L. Flattening the last two dimensions results in a 2-dimensional dataset with the first dimension of size k and the second dimension of size n*L. Then, a new dimension of size 1 is added between the first dimension k and the second dimension n*L. Finally, the flattened result is mapped to the final external features using a linear layer. This mapping process transforms the n*L dimension to the size d_model.

[0093] Step 4: Input the prior features and intrinsic features into the prior feature-guided cross-attention module, and perform temporal dimension fusion with the prior features as a guide to generate fused prior intrinsic features.

[0094] In this embodiment, prior features and intrinsic features are input into the attention module that guides the cross-reference of prior features. The prior features are used as queries, and the intrinsic features are used as keys and values. The similarity between the prior features and intrinsic features is calculated through the attention mechanism. The prior features are then integrated into the intrinsic features to obtain the intrinsic features that are fused with the prior features.

[0095] Among them, the attention module guided by prior features, such as Figure 5 As shown, it's worth noting that cross-attention has been widely used in multimodal learning to capture the dependencies between adaptive tokens across different modalities. This step involves two cross-attention calculations. The first is a cross-attention calculation between prior features and intrinsic features, essentially a sequence-level fusion of learned global features between prior and historical tracks.

[0096] Specifically, the implementation process of the attention module guided by prior features is as follows:

[0097] The prior features obtained in step 3 As a query, the intrinsic characteristics of historical tracks

[0098] As the key, the value is input to the output of the cross-attention layer. This can be expressed by the formula:

[0099]

[0100] Step 5: Input the intrinsic and extrinsic features of the fusion prior into the extrinsic cross-attention module to perform variable-level feature fusion and generate the final fused features.

[0101] Among them, the external cross-attention module, such as Figure 5 As shown, the intrinsic and extrinsic features of the fused prior features are input into the extrinsic cross-attention module. The intrinsic features of the fused prior features are used as the query, and the extrinsic features are used as the key and value. The similarity between the intrinsic and extrinsic features of the fused prior features is calculated through cross-attention to obtain the final fused features.

[0102] Specifically, will As a query, external features As the key and value, they undergo another cross-attention mechanism, followed by layer normalization, a feedforward layer, and another layer normalization to obtain the final fused feature. This can be expressed by the formula:

[0103]

[0104] Step 6: Concatenate the final fused features with the historical track sequence, and generate the future 4D track prediction result in one step through linear mapping.

[0105] The specific implementation process is as follows:

[0106] This method uses a one-step prediction approach to generate the final predicted track, avoiding the cumulative error caused by step-by-step prediction. Error accumulation can be specifically explained as follows: to predict the track t steps ahead, and the actual historical track is X... Ture =[p1,p2,…,p L The step-by-step prediction method is first based on [p1, p2, ..., p... L Output p′ L+1 Then, based on the sequence [p2,…,p L ,p′ L+1 ], predict output p′ L+2 Predict step by step until p′ L+t Obviously, using the new predicted data p′ as input to the model during this process will cause error accumulation. However, the one-step generative prediction method in this paper can adjust the dimensions through fully connected layers and output complete time-series data [p′] based on the label sequence. L+1 ,…,p′ L+t To avoid error accumulation.

[0107] Finally, the final output feature O of the attention module, which is guided by multiple layers of prior features, is... finall With Pin By splicing Then, perform a Flatten operation on the last two dimensions to obtain the dimensions as follows: The tensor outputs a one-step generative predicted trajectory sequence. This avoids the error accumulation that occurs in traditional step-by-step prediction methods. The specific one-step prediction process can be expressed by the formula:

[0108]

[0109] Understandably, steps 2-6 in this embodiment constitute a 4D trajectory prediction model. This embodiment also provides a training method for the 4D trajectory prediction model, the process of which is as follows:

[0110] First, the batch size and number of training epochs are set to control the model training effect, and MAE and MSE are used as evaluation metrics for the model results:

[0111]

[0112] Where Y represents the predicted value. Represents the actual value.

[0113] The 4D trajectory prediction model based on prior trajectory guidance and multivariate time lag dynamic adjustment is set with a batch size of 32, an epoch of 30, a learning rate of 0.0001, and the Adam optimizer is used to progressively adjust the learning rate. MSE is used as the loss to train the model.

[0114] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A 4D trajectory prediction method based on prior trajectory guidance and multivariate time lag dynamic adjustment, characterized in that, Includes the following steps: Step 1: Preprocess the real historical flight data and meteorological data; Step 2: Extract prior track data of fixed navigation points from the flight navigation database and input it into the prior feature extraction module to generate prior features; Simultaneously, the intrinsic variables of the preprocessed real historical flight data are input into the intrinsic feature extraction module to obtain intrinsic features, and the historical flight track processed by the intrinsic feature extraction module is obtained. Step 3: Input the exogenous variables of the preprocessed real historical flight data into the adaptive time-delay convolution exogenous feature encoding module to generate exogenous features; Step 4: Input the prior features and intrinsic features into the prior feature-guided cross-attention module, and perform temporal fusion with the prior features as a guide to generate the fused prior intrinsic features; Step 5: Input the intrinsic and extrinsic features of the fusion prior into the extrinsic cross-attention module to perform variable-level feature fusion and generate the final fused features; Step 6: Concatenate the final fused features with the historical trajectory sequence after the endogenous feature extraction module, and generate the future 4D trajectory prediction result in one step through linear mapping.

2. The 4D trajectory prediction method based on prior trajectory guidance and multivariate time lag dynamic adjustment according to claim 1, characterized in that, The preprocessing methods include: outlier handling, trackpoint interpolation, temporal feature standardization, and data alignment.

3. The 4D trajectory prediction method based on prior trajectory guidance and multivariate time lag dynamic adjustment according to claim 1, characterized in that, In step 2, the operations of the prior feature extraction module include: Perform a patching operation on the prior track data to divide it into segments of length P; The fragment is linearly mapped to a high-dimensional feature and concatenated with a learnable global token. Global trajectory features are extracted using a self-attention mechanism, and prior features are output.

4. The 4D trajectory prediction method based on prior trajectory guidance and multivariate time lag dynamic adjustment according to claim 3, characterized in that, The endogenous variables include longitude, latitude, altitude, and time.

5. The 4D trajectory prediction method based on prior trajectory guidance and multivariate time lag dynamic adjustment according to claim 4, characterized in that, The external variables include flight number, temperature, humidity, and wind speed.

6. The 4D trajectory prediction method based on prior trajectory guidance and multivariate time lag dynamic adjustment according to claim 4, characterized in that, In step 2, the operation of the intrinsic feature extraction module is the same as that of the prior feature extraction module. The input of the intrinsic feature extraction module is to stitch together multiple historical flight tracks of real historical flight data into a sequence by splicing.

7. The 4D trajectory prediction method based on prior trajectory guidance and multivariate time lag dynamic adjustment according to claim 5, characterized in that, In step 2, the operation of the adaptive time-delay convolutional external feature encoding module includes: First convolutional layer: First, adjust the dimensionality of the n historical track data with k external variables to [value]. Intermediate results were obtained using the grouped Conv1d method, and their dimensions were adjusted to obtain periodic features. Perform a Gumbel-Softmax operation on the obtained periodic feature C, and set hard=True to convert the output into a strict one-hot vector; The second convolutional layer uses a time-lag window kernel that employs a strictly one-hot vector as the second convolutional layer. The second convolutional layer first processes X... ex Padding is performed only at the end of the L-axis, using a moving average method. The padding length is l-1, adjusted for dimension variation. by Grouped convolution is used for the convolution kernel, for X padex After performing the convolution operation, adjust the dimensions to obtain That is, to perform time-lag alignment on k external variables of n historical tracks; Finally The last two dimensions are flattened into one dimension, and a linear layer is used to map the flattened dimension size from n*L to the hidden representation dimension size d_model to obtain the final external feature.

8. The 4D trajectory prediction method based on prior trajectory guidance and multivariate time lag dynamic adjustment according to claim 7, characterized in that, In step 4, the operation of the prior feature-guided cross-attention module includes: Based on prior features As a query, internal features As Key and Value; By calculating similarity through cross-attention, the physical constraints of prior tracks are incorporated into historical track features to obtain intrinsic features that fuse prior information.

9. The 4D trajectory prediction method based on prior trajectory guidance and multivariate time lag dynamic adjustment according to claim 8, characterized in that, In step 5, the operation of the external cross-attention module includes: Integrate prior intrinsic features As a query, external features As the key and value, they undergo another cross-attention mechanism, followed by layer normalization, a feedforward layer, and another layer normalization to obtain the final fused feature.

10. The 4D trajectory prediction method based on prior trajectory guidance and multivariate time lag dynamic adjustment according to claim 8, characterized in that, Step 6 employs a one-step generative prediction method, specifically including: Final output feature O finall Compared with historical tracks after endogenous feature extraction module By splicing After flattening the last two dimensions, a linear layer mapping operation is used to further obtain the dimension as follows. The tensor is the predicted trajectory sequence.

11. The 4D trajectory prediction method based on prior trajectory guidance and multivariate time lag dynamic adjustment according to claim 6, characterized in that, The endogenous feature extraction module additionally incorporates temporal encoding: The time characteristics of historical flight tracks are broken down into year, month, day, hour, minute, and second components; The location information is supplemented by encoding through the Embedding layer and then superimposed onto the track features.

Images