Uncertain Flight Trajectory Prediction Model and Method Incorporating Flight Dynamics Knowledge

Abstract

The present invention proposes an uncertain flight trajectory prediction model and method integrating flight dynamics knowledge. The prediction model includes a trajectory context feature encoding module, a pre-trained LLM module, and a decoding module based on a flight dynamics model. The present invention first proposes introducing a pre-trained large language model into the field of flight trajectory prediction, combining flight state features with a flight dynamics model, and designing a dedicated encoding-decoding process for the LLM. This integration enables FlightLLM to fully leverage the advantages of the LLM in temporal data reasoning, use the temporal reasoning ability of the LLM to model complex flight dynamics, and at the same time constrain the physical rationality of the trajectory through the dynamics model, thereby effectively representing the complex dynamic characteristics in flight trajectory prediction. Moreover, for the FlightLLM architecture, the present invention designs a multi-expert adapter, inserts a trainable adapter into the pre-trained LLM, and realizes the improvement of prediction accuracy and uncertainty quantification through ensemble learning and random masking.

Description

Technical Field

[0001] The present invention relates to the technical field of flight trajectory prediction, and particularly to an uncertain flight trajectory prediction model and method integrating flight dynamics knowledge. Background Art

[0002] Flight Trajectory Prediction (FTP), as the core technology of Air Traffic Management (ATM), its core goal is to accurately estimate the future flight state by analyzing the historical observation data of aircraft (including spatio-temporal information such as position and speed), and it has important application value in key scenarios such as flight scheduling, conflict detection, and airspace traffic management. Existing technologies can be divided into two categories: physical model-driven and data-driven methods, each with significant technical characteristics and limitations:

[0003] Traditional physical model-driven FTP methods are based on flight dynamics equations, and construct state evolution models by integrating aerodynamic / kinematic models and state estimation algorithms such as Kalman filter. Some improved studies improve the prediction accuracy by introducing prior information such as flight intention recognition or Estimated Time of Arrival (ETA). However, its effectiveness is restricted by two essential defects: (1) The model mismatch phenomenon causes the prediction error to show a cumulative effect in multi-step iterations; (2) The flight dynamics system has strong nonlinear and complex environment coupling characteristics, and there are engineering challenges such as difficult parameter identification and high computational complexity in constructing a high-precision physical model, resulting in the prediction accuracy being difficult to meet the actual application requirements.

[0004] With the development of machine learning technology, data-driven FTP methods with a multi-step multi-variable time series prediction as the modeling framework have shown significant advantages. They directly extract spatio-temporal features from offline data through machine learning without relying on explicit motion models. Initial studies mainly used RNN and its variants (LSTM / GRU) to capture temporal dependencies, and used means such as Monte Carlo Dropout and Bayesian neural networks for uncertainty quantification. The current technological frontier presents three evolution directions: (1) Models based on the Transformer architecture use self-attention mechanisms to enhance the ability to capture long-range features; (2) Combining preprocessing techniques such as wavelet transform and trajectory binarization to optimize feature representation; (3) Trying to embed flight dynamics constraints into the neural network structure design. Nevertheless, there is still room for improvement in the dynamic environment adaptability, physical interpretability, and generalization ability in complex airspace scenarios of existing methods.

[0005] Large language models (LLMs) represented by GPT / LLaMA, with their ultra-large-scale parameters and the advantages of the Transformer architecture, show unique potential in long-sequence modeling and transfer learning. Through domain-adaptive training, LLMs are expected to achieve dynamic correlation modeling of flight state information to improve FTP performance. However, practical applications face three challenges: (1) insufficient verification of the matching degree between the highly dynamic characteristics of flight trajectories and the time-series modeling ability of LLMs; (2) the lack of a reliable uncertainty quantification mechanism in existing LLM frameworks; (3) the effective architecture for the integration of flight dynamics prior knowledge and data-driven models has not been formed. Summary of the Invention

[0006] Aiming at the technical problems existing in the existing data-driven FTP methods, the present invention proposes an uncertainty flight trajectory prediction model and method integrating flight dynamics knowledge, abbreviated as FlightLLM. This method combines a flight dynamics model with a pre-trained LLM. It not only uses the flight dynamics model to provide the coupling relationship and evolution constraints between state components based on physical laws to improve prediction accuracy and enhance interpretability, but also uses the powerful time-series and semantic reasoning capabilities of the LLM to effectively model complex flight dynamics.

[0007] The technical solution of the present invention is as follows:

[0008] An uncertainty flight trajectory prediction model integrating flight dynamics knowledge, including a trajectory context feature encoding module, a pre-trained LLM module, and a decoding module based on a flight dynamics model;

[0009] The trajectory context feature encoding module is used to map the historical observation sequence of the flight trajectory into a high-dimensional vector compatible with the pre-trained LLM module ; is the time length of the historical observation sequence, is the set number of repetitions;

[0010] The pre-trained LLM module includes pre-trained LLM layers and expert adapter layers between adjacent pre-trained LLM layers; The pre-trained LLM layers directly call existing large language models. During training, the parameters of the pre-trained LLM layers are frozen, and only the parameters of the expert adapter layers are optimized; the input of the pre-trained LLM module is , and the output is a high-dimensional feature sequence ;

[0011] The decoding module includes network network , a flight dynamics model solver module, and a compensation module;

[0012] The network The high-dimensional feature sequence output by the pre-trained LLM module The last features are mapped to the predicted overload instruction sequence ;

[0013] The flight dynamics model solving module uses the final observed state of the flight trajectory historical observation sequence as the initial value, and based on the predicted overload instruction sequence , calculates and obtains to the coarse-grained flight trajectory sequence at the moment ;

[0014] The network uses the coarse-grained flight trajectory sequence and the last features in the high-dimensional feature sequence as inputs and outputs the compensation term ;

[0015] The compensation module supplements the compensation term into the coarse-grained flight trajectory sequence to obtain the final prediction result .

[0016] Furthermore, the trajectory context feature encoding module includes an observation trajectory encoding network and a differential position encoding network ;

[0017] The observation trajectory encoding network processes the input flight trajectory historical observation sequence to generate a high-dimensional feature vector ;

[0018] The differential position encoding network processes the input differential position sequence to generate a high-dimensional query feature vector ; The differential position sequence is obtained by performing a differential calculation on the flight trajectory historical observation sequence ;

[0019] The high-dimensional query feature vector is copied times and then concatenated with the high-dimensional feature vector to obtain the input feature of the pre-trained LLM module .

[0020] Furthermore, the observation trajectory encoding network and the differential position encoding network are each composed of a GRU layer and an MLP layer. The GRU layer captures the temporal features of the input sequence, and the MLP layer further fits it.

[0021] Furthermore, the network is composed of an MLP layer; the network is composed of an MLP layer and a GRU layer, and after being respectively fitted through their respective corresponding MLP layers, are combined and output the compensation term through the GRU layer .

[0022] The present invention also proposes an uncertainty flight trajectory prediction method that integrates flight dynamics knowledge based on the above prediction model, including the following steps:

[0023] Step 1: Obtain the historical observation sequence of the flight trajectory for prediction:

[0024]

[0025] where is the time length of the historical observation sequence, represents the moment , and , is the observation state at the moment .

[0026] Step 2: Input the historical observation sequence of the flight trajectory obtained in Step 1 into the trained prediction model to predict the future flight trajectory sequence:

[0027]

[0028] where is the time length of the prediction sequence, is the prediction state at the moment , is the prediction state at the moment .

[0029] Furthermore, in Step 1, , where is the three-axis coordinate component of the flight trajectory point in the ground coordinate system at the moment , is the three-axis velocity component of the flight trajectory point in the ground coordinate system at the moment .

[0030] Furthermore, the training process of the prediction model includes:

[0031] Obtain several segments of flight trajectory data as the training sample dataset. For each training sample, it is divided into an observed trajectory and a future trajectory. The training samples are input into the prediction model to encode the observed trajectory in the trajectory context feature encoding module parameters and the differential position encoding network parameters, the parameters of the expert adapter layer in the pre-trained LLM module, and the network in the decoding module parameters and the network parameters are trained.

[0032] Furthermore, during the training process of the prediction model, the pre-trained LLM module adopts an uncertainty quantification mode: when each training sample input is trained, a random expert adapter layer is activated for training, and the remaining expert adapter layers are not activated; during actual prediction, the pre-trained LLM module adopts an accuracy mode: all expert adapter layers are activated, and the knowledge of the multi-expert adapter layer network is aggregated through ensemble learning techniques to optimize the prediction results.

[0033] Furthermore, during the training process of the prediction model, the Huber loss function is adopted to enhance the robustness of the model to noise and outliers.

[0034] Beneficial effects:

[0035] The present invention has the following effects:

[0036] 1. First, a large language model-driven flight trajectory prediction method FlightLLM is proposed.

[0037] The present invention first proposes to introduce a pre-trained large language model into the field of flight trajectory prediction, combines flight state features with the flight dynamics model, and designs a dedicated encoding-decoding process for the LLM. This integration enables FlightLLM to fully utilize the advantages of the LLM in temporal data reasoning, use the temporal reasoning ability of the LLM to model complex flight dynamics, and at the same time constrain the physical rationality of the trajectory through the dynamics model, thereby effectively representing the complex dynamic characteristics in flight trajectory prediction.

[0038] 2. A customized multi-expert adapter for uncertainty quantification is proposed.

[0039] For the FlightLLM architecture, the present invention designs a multi-expert adapter, inserts a trainable adapter into the pre-trained LLM, and realizes the improvement of prediction accuracy and uncertainty quantification through ensemble learning and random masking. This adapter can provide two inference modes:

[0040] (1) High-precision mode. The knowledge of multiple expert networks is aggregated through ensemble learning techniques to optimize the prediction results and significantly improve the accuracy.

[0041] (2) Uncertainty quantification mode. By evaluating the prediction uncertainty and combining the potential changes in flight trajectories, more reliable predictions are provided. This mode is particularly important in the aviation field, and its uncertainty assessment is crucial for safety decisions.

[0042] Experiments based on the real trajectory dataset of commercial flights around the airport show that FlightLLM significantly outperforms existing models in prediction accuracy and also demonstrates effective uncertainty quantification ability, becoming a potential tool to improve the safety and efficiency of air traffic control.

[0043] Additional aspects and advantages of the present invention will be given in part in the following description, will become apparent in part from the following description, or will be understood through the practice of the present invention. Brief Description of the Drawings

[0044] The above and / or additional aspects and advantages of the present invention will become apparent and be readily understood from the description of the embodiments in conjunction with the following drawings, in which:

[0045] Figure 1 : Schematic diagram of flight dynamics model parameters;

[0046] Figure 2 : Overall architecture of FlightLLM;

[0047] Among them: (I) Trajectory context feature encoding module: Encodes the observed flight trajectories into high-dimensional features understandable by the LLM; (II) Pre-trained LLM module with multiple expert adapters: Serves as the core architecture for capturing deep temporal features; (III) Decoding module based on the flight dynamics model: Decodes the high-dimensional features into flight overload sequences through the flight dynamics model to predict the state and performs fine-grained compensation on the prediction results;

[0048] Figure 3 : Multi-expert adapter architecture;

[0049] Among them: The left frame is: Multiple expert adapters are embedded in series between each attention layer of the LLM for parameter fine-tuning;

[0050] The upper right frame is: In the training phase and the sampling inference mode for uncertainty quantification, multiple expert adapters are generated by splitting the complete network, and one of them is randomly selected in each calculation;

[0051] The lower right frame is: In the integrated inference mode for improving prediction accuracy, multiple expert adapters will be combined for joint calculation;

[0052] Figure 4 : Schematic diagram of the position components of the predicted trajectories of each method in the xyz directions;

[0053] Among them: (a) Low-maneuver ascent trajectory, (b) Low-maneuver descent trajectory, (c) High-maneuver ascent trajectory, (d) High-maneuver descent trajectory;

[0054] Figure 5 : The predicted mean and 3σ upper and lower bound ranges of the sampled trajectories in the xyz directions based on multiple sets of Monte Carlo predictions;

[0055] The orange solid line in the figure is the predicted mean, the orange shaded area is the mean ± 3σ confidence interval, and the black solid line is the true trajectory;

[0056] Figure 6 : Comparison of the position prediction performance results of each method at different prediction steps;

[0057] Figure 7 : Comparison results of each method at different flight stages (15-step prediction);

[0058] Figure 8 : Comparison of the speed prediction performance results of each method (15-step prediction);

[0059] Figure 9 : Comparison of experimental results under different ablation scenarios (15-step prediction). Detailed implementation manners

[0060] The embodiments of the present invention are described in detail below. The embodiments are exemplary and are intended to explain the present invention and should not be construed as limiting the present invention.

[0061] (1) Embodiment:

[0062] This embodiment aims at short-term FTP tasks, aiming to predict the flight states in the future for a period of time, such as position and speed, based on historical trajectory data, so as to improve the real-time safety management performance in high-density airspaces such as airports. For this purpose, an uncertainty flight trajectory prediction method FlightLLM that integrates flight dynamics knowledge is proposed. Figure 2 The FlightLLM framework proposed in this embodiment is shown as follows: Using a pre-trained LLM as the backbone, integrating the prior knowledge of the flight dynamics model and offline data, predicting the future trajectory in an efficient non-autoregressive form, and supporting uncertainty quantification. This method combines deep learning technology and domain knowledge to provide a robust and scalable solution for FTP.

[0063] The uncertainty flight trajectory prediction method that integrates flight dynamics knowledge in this embodiment specifically includes the following steps:

[0064] Step 1: The essence of flight trajectory prediction is to infer the evolution law of future flight states through historical observation data. Therefore, first obtain the historical observation sequence of the flight trajectory for prediction:

[0065]

[0066] wherein is the time length of the historical observation sequence, represents the moment and , is the observation state at the moment , wherein is, at the moment , the three-axis coordinate components of the flight trajectory point in the ground coordinate system, is, at the moment , the three-axis velocity components of the flight trajectory point in the ground coordinate system;

[0067] Step 2: Input the flight trajectory historical observation sequence obtained in Step 1 into the trained prediction model to predict the future flight trajectory sequence:

[0068]

[0069] wherein is the time length of the prediction sequence, is the prediction state at the moment , is the prediction state at the moment , and the meanings of the components in the prediction state are the same as those in the observation state.

[0070] As Figure 2 shown, the prediction model includes a trajectory context feature encoding module, a pre-trained LLM module with multiple expert adapters, and a decoding module based on the flight dynamics model.

[0071] Trajectory context feature encoding module:

[0072] The trajectory context feature encoding module is used to map the flight trajectory historical observation sequence into a high-dimensional vector compatible with the large language model (LLM). As Figure 2 shown in the first part, the trajectory context feature encoding module includes an observation trajectory encoding network and a differential position encoding network ; both the observation trajectory encoding network and the differential position encoding network are respectively composed of a GRU layer and an MLP layer. The GRU layer is used to capture the temporal features of the input sequence, and the MLP layer is further used for fitting to generate a high-dimensional feature vector.

[0073] wherein the flight trajectory historical observation sequence is input into the observation trajectory encoding network , first through the GRU layer to capture its temporal characteristics, and then through the MLP layer for further fitting to generate a high-dimensional feature vector , is the input dimension of the pre-trained LLM module.

[0074] In addition, considering that position change is a key indicator of flight dynamics, this embodiment also uses the flight trajectory historical observation sequence Perform differential calculation to obtain the differential position sequence ; The differential position sequence Input the differential position encoding network , also through the GRU layer to capture its time series features, and then further fit through the MLP layer to generate a high-dimensional query feature vector ; The differential position sequence captures the motion information at different time scales, which is crucial to understand the motion evolution. copy Secondary and high-dimensional feature vectors Splicing to obtain the complete input features of the pre-trained LLM module The number of copies here This is the time length of the forecast sequence.

[0075] Pre-trained LLM modules with multiple expert adapters:

[0076] The pre-trained LLM module includes There are two pre-trained LLM layers and expert adapter layers between adjacent pre-trained LLM layers. The pre-trained LLM layer directly calls the existing commercial large language model. During training, the pre-trained LLM layer is frozen and only the expert adapter layer is optimized. The input of the pre-trained LLM module is , output high-dimensional feature sequence .

[0077] Decoding module based on flight dynamics model:

[0078] like Figure 2 As shown in the third part of the paper, the high-dimensional feature sequence output by the LLM module is The last of Features Via the network Mapped to predicted overload instruction sequence , the network It consists of one MLP layer.

[0079] The last observation state of the flight trajectory historical observation sequence is the initial value, based on the predicted overload instruction sequence , calculated by the flight dynamics model to coarse-grained flight trajectory sequence at a moment Each predicted state in the coarse-grained flight trajectory sequence is also composed of three-axis coordinate components and three-axis velocity components in the ground coordinate system.

[0080] As Figure 1 shown, the flight dynamics model in this embodiment adopts a three-degree-of-freedom particle model. According to the overload vector, the three-axis coordinate components and velocity components of the particle can be solved, so that by giving the initial state and combining the predicted overload command sequence, the coarse-grained future flight trajectory sequence can be recursively predicted through the three-degree-of-freedom particle model.

[0081] Also because the three-degree-of-freedom particle model adopted is a simplified approximation of the real flight process and there are large errors in its prediction, the coarse prediction result is further refined through a compensation mechanism. Specifically, the coarse-grained flight trajectory sequence and the high-dimensional feature sequence output by the LLM module the last feature are jointly input into the network to obtain a compensation term The compensation term is supplemented into the coarse-grained flight trajectory sequence to obtain the final prediction result:

[0082]

[0083] where the network is composed of two MLP layers and one GRU layer. and After being respectively fitted through one MLP layer and combined, the compensation term is output through the GRU layer.

[0084] The training process of the prediction model includes:

[0085] Obtain several segments of flight trajectory data as the training sample data set. For each training sample, it is divided into an observed trajectory and a future trajectory. The training sample is input into the prediction model to train the parameters of the observed trajectory encoding network in the trajectory context feature encoding module and the differential position encoding network , the parameters of the expert adapter layer in the pre-trained LLM module, the network in the decoding module, and the network parameters, where the Huber loss function is used to enhance the robustness of the model to noise and outliers.

[0086] During training, the pre-trained LLM module adopts an uncertainty quantization mode. That is, when each training sample input is used for training, one expert adapter layer is randomly activated for training, and the remaining expert adapter layers are not activated.

[0087] During actual prediction, the pre-trained LLM module adopts an accuracy mode, activating all expert adapter layers, and aggregating the knowledge of multiple expert adapter layer networks through ensemble learning techniques to optimize the prediction results, significantly improving the accuracy.

[0088] (2) Case study verification:

[0089] Case study settings:

[0090] Dataset: In the case study, we use the publicly available flight trajectory dataset provided by the literature (Gariel, M.; Srivastava, A. N.; and Feron, E. 2011. Trajectory clustering and an application to airspace monitoring. IEEE Transactions on Intelligent Transportation Systems, 12(4): 1511–1524.). The trajectory point interval is set to 20 seconds, and the total trajectory length is set to 480 seconds. For the preprocessed data resampled at 4-second intervals, every 5th trajectory point is selected to create trajectory segments of length 24 steps. Trajectories shorter than 24 steps are excluded, and trajectories of 24 steps or longer are divided into consecutive, non-overlapping 24-step segments, with the remaining points discarded. The final obtained trajectory segments use the first 9 steps as the model input and the last 15 steps as the ground truth labels for prediction. Finally, we obtained 13,000 trajectory segments, of which 10,000 are used for training, 2,000 for validation, and 1,000 for testing.

[0091] In the case study verification, the comparative experiments selected two time-series deep learning models commonly used for flight trajectory prediction (FTP): the autoregressive model and the parallel prediction model.

[0092] The autoregressive model is constructed by bidirectional LSTM (Bi-LSTM) and bidirectional gated recurrent unit (Bi-GRU), and uses recursion to predict future trajectories.

[0093] The parallel prediction model is based on the encoder-decoder structure of Transformer and the LSTM+Attention architecture that combines the attention mechanism, and adopts a sequence-to-sequence (seq2seq) computing architecture.

[0094] Evaluation metric settings: To evaluate the accuracy of position and speed estimation, this embodiment adopts three core metrics:

[0095] 1. Root Mean Squared Error (RMSE)

[0096]

[0097] 2. Mean Absolute Error (MAE)

[0098]

[0099] 3. Mean Distance Error (MDE)

[0100]

[0101] Prediction accuracy analysis

[0102] To comprehensively evaluate the prediction accuracy of FlightLLM proposed in this embodiment, the adapter is configured to the integrated inference mode and compared with other comparison algorithms at three different prediction lengths of 3, 9, and 15. For fairness, the sequence-to-sequence methods are retrained for each prediction length.

[0103] Figure 6 The statistical results of the position prediction performance of each method are shown. In the case of short-term prediction (prediction length 3), the dynamic characteristics and complexity of the flight trajectory sequence are not fully captured, making the advantages of the sequence-to-sequence methods in dealing with long-term dependencies and trends not obvious. In the case of long-term prediction, the sequence-to-sequence methods can better handle the long-term dependencies and complex dynamic characteristics of the trajectory through end-to-end learning of the long-term prediction process. The complex maneuvers of the flight target will cause significant error accumulation in autoregressive methods such as Bi-GRU and Bi-LSTM. Compared with other methods, FlightLLM provides excellent long-term prediction performance while ensuring short-term prediction accuracy, relying on the integration of prior model knowledge and the powerful time series feature capture ability of the pre-trained large language model. Another advantage of integrating the prior flight dynamics model is reflected in the improvement of speed prediction performance, as Figure 8 shown in the speed statistical results of the 15-step prediction length. Through inference based on the prior dynamics model, FlightLLM effectively captures the coupling relationship between position and speed, thus performing outstandingly in speed prediction.

[0104] Figure 7Further shows the prediction performance of each method during takeoff and landing phases. For the problem of flight trajectory prediction near airports, the prediction of takeoff trajectories is significantly more difficult than that of landing trajectories. During the landing phase, the aircraft usually has a clear destination, its movement is more geographically sensitive and easier to learn from offline data. However, the takeoff process of the aircraft has more diverse variations, resulting in lower overall prediction accuracy. Obviously, the powerful data learning ability of FlightLLM enables it to capture the movement characteristics of the target at different flight stages, thus achieving accurate predictions for both takeoff and landing phases, and the prediction accuracy is better than other methods.

[0105] In addition, in Figure 4 it shows the comparison of trajectory prediction results between FlightLLM and other methods (such as Bi-LSTM, Transformer, etc.) in different maneuver scenarios through four groups of subgraphs:

[0106] Low-maneuver scenarios (a, b): The movement of the aircraft is gentle and the trajectory changes little. The predicted trajectory (dashed line) of FlightLLM (red) highly coincides with the true trajectory (black solid line), while other methods (such as blue, green) gradually deviate during long-term prediction;

[0107] High-maneuver scenarios (c, d): The movement of the aircraft is complex and the trajectory changes violently. By combining dynamic priors, FlightLLM significantly suppresses the error accumulation of traditional methods (such as Bi-GRU); in the prediction of the z-direction (altitude change), the fluctuation range of FlightLLM is closer to the true value;

[0108] Obviously, FlightLLM achieves the best prediction performance in both low-maneuver and high-maneuver situations. In contrast, the autoregressive BLSTM and BGRU show significant prediction biases due to error accumulation. Although LSTM+Attention and Transformer can better learn the changing rules of flight movements, their overall prediction performance is still inferior to FlightLLM. More importantly, by learning overload instructions and performing trajectory prediction based on the dynamic model, the predicted trajectory of FlightLLM is constrained by flight dynamics, ensuring the physical rationality of the predicted trajectory while improving the accuracy. Figure 4 Verifies the advantages of FlightLLM in complex maneuver adaptability and long-term prediction stability, and its physical constraint mechanism (dynamic model) effectively improves the trajectory rationality.

[0109] Prediction Uncertainty Analysis

[0110] To evaluate the ability of FlightLLM to quantify prediction uncertainty, this embodiment also tests its sampling inference scheme. Specifically, 20 Monte Carlo predictions are performed on the test data, and the mean and variance are respectively statistically analyzed. Figure 5The three-dimensional visualization shows the sampling trajectories and describes the predicted means and ±3σ bounds (σ is the standard deviation) in the x, y, and z directions. Obviously, the accuracy is relatively high in the initial stage of prediction. The sampling trajectories predicted by FlightLLM have a small dispersion, indicating low prediction uncertainty. Therefore Figure 5 In the early and middle stages, the sampling trajectories almost overlap, and the corresponding dispersions in the x, y, and z directions are small. As the prediction time extends, the climbing of the aircraft is usually accompanied by changes in the motion pattern, resulting in a decrease in the prediction accuracy in the later stage. At this time, the sampling trajectories of FlightLLM have a higher dispersion and a larger variance, enabling it to output diverse prediction results. Therefore, FlightLLM can effectively measure the prediction uncertainty through the changes in dispersion or variance, evaluate the prediction risk, and identify abnormal motion patterns in the trajectories.

[0111] Figure 5 Shows the performance of FlightLLM in uncertainty quantification:

[0112] Sampling trajectories: By randomly activating different expert adapters (multi-expert structure), multiple possible trajectories are generated.

[0113] Confidence intervals: In the initial stage (the first 5 steps), the prediction uncertainty (σ value) is small, and the confidence intervals are narrow, indicating that the model has a high confidence in short-term predictions; as the prediction step increases (the last 10 steps), the confidence intervals gradually expand, reflecting the natural growth of uncertainty.

[0114] Key findings:

[0115] In the high-maneuver stage (such as rapid climbing in the z direction), the width of the confidence interval is positively correlated with the amplitude of the trajectory mutation, indicating that the model can perceive the maneuver risk.

[0116] The true trajectory (black) is always within the 3σ interval, proving the reliability of the uncertainty estimation.

[0117] Analysis of the effectiveness of each module

[0118] The effectiveness of each module is evaluated by disabling specific modules in FlightLLM. Specifically, three ablation scenarios are set:

[0119] 1. Remove the multi-expert adapters in FlightLLM, making the LLM act as a single backbone network without fine-tuning through adapters;

[0120] 2. Exclude the output of the dynamics model and directly use the output of a single-layer MLP with 128 nodes as the input to predict the trajectory;

[0121] 3. Disable the differential query calculation, that is, do not use to calculate but set it as a learnable parameter vector.

[0122] As shown Figure 9 In the figure, the prediction performance of these three ablation scenarios is tested with a 15-step prediction length. Obviously, the absence of any module will lead to a decline in the prediction performance of FlightLLM. First, the absence of the multi-expert adapter has the greatest impact on the overall performance, as it enhances the model's ability to understand and learn complex features through fine-tuning, thus improving the adaptability to complex tasks and real data. The absence of the dynamics model will significantly affect the speed prediction performance, as it effectively captures and utilizes the coupling relationship between speed and position. In addition, the differential query also has a significant impact, as it introduces dynamic change information at different time scales beyond the original speed features, capturing the long-term motion patterns in the flight trajectory data.

[0123] In summary, this embodiment proposes an innovative flight trajectory prediction method FlightLLM. This model uses a pre-trained large language model (LLM) as the core architecture and realizes efficient short-term flight trajectory prediction by integrating prior flight dynamics model knowledge. We designed a new type of multi-expert adapter that supports the quantification of prediction uncertainty while improving the prediction accuracy. A large number of experiments show that FlightLLM can achieve excellent prediction accuracy under different maneuver intensities and effectively quantify the prediction uncertainty, thus ensuring both high accuracy and high reliability in trajectory prediction. Generally speaking, FlightLLM provides a promising technical tool for improving air traffic control safety and operation efficiency by providing accurate and reliable trajectory predictions.

[0124] Although the embodiments of the present invention have been shown and described above, it can be understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those of ordinary skill in the art can make changes, modifications, substitutions, and variations to the above embodiments within the scope of the present invention without departing from the principles and purposes of the present invention.

Claims

1. An uncertain flight trajectory prediction model integrating flight dynamics knowledge, characterized by: It includes trajectory context feature encoding module, pre-trained LLM module and decoding module based on flight dynamics model; The trajectory context feature encoding module is used to map the flight trajectory historical observation sequence into a high-dimensional vector compatible with the pre-trained LLM module. ; is the time length of the historical observation sequence, is the set number of repetitions; The pre-trained LLM module includes Layer pre-trained LLM layers and expert adapter layers between adjacent pre-trained LLM layers; The pre-trained LLM layer directly calls the existing large language model, freezes the pre-trained LLM layer parameters during training, and only optimizes the expert adapter layer parameters; the input of the pre-trained LLM module is , output high-dimensional feature sequence ; The decoding module includes a network ,network , flight dynamics model solving module and compensation module; The network The high-dimensional feature sequence output by the pre-trained LLM module The last of Features Mapped to predicted overload instruction sequence ; The flight dynamics model solving module uses the last observation state of the flight trajectory historical observation sequence is the initial value, based on the predicted overload instruction sequence , calculated to Coarse-grained flight trajectory sequence at each moment ; The network Coarse-grained flight trajectory sequence And high-dimensional feature sequences The last of Features is the input and output compensation term ; The compensation module will compensate the Add the coarse-grained flight trajectory sequence to obtain the final prediction result .

2. The uncertain flight trajectory prediction model integrating flight dynamics knowledge according to claim 1 is characterized in that: The trajectory context feature encoding module includes an observation trajectory encoding network and differential position encoding network ; The observation trajectory encoding network The input flight trajectory history observation sequence Processing to generate high-dimensional feature vectors ; The differential position encoding network The differential position sequence of the input Processing to generate high-dimensional query feature vector The differential position sequence By observing the flight trajectory history Perform differential calculation to obtain; The high-dimensional query feature vector copy Secondary and high-dimensional feature vectors Splicing to obtain the input features of the pre-trained LLM module .

3. The uncertain flight trajectory prediction model integrating flight dynamics knowledge according to claim 2 is characterized in that: The observation trajectory encoding network and the differential position encoding network They are composed of GRU layer and MLP layer respectively. The temporal features of the input sequence are captured by the GRU layer and further fitted by the MLP layer.

4. The uncertain flight trajectory prediction model integrating flight dynamics knowledge according to claim 1 is characterized in that: The network The network consists of MLP layers; It consists of MLP layer and GRU layer. and After each is fitted through its corresponding MLP layer, the combination is output through the GRU layer to compensate .

5. The uncertain flight trajectory prediction method based on the prediction model described in any one of claims 1 to 4 and integrating flight dynamics knowledge, characterized in that: The following steps are involved: Step 1: Get the historical observation sequence of the flight trajectory for prediction: in is the time length of the historical observation sequence, Indicates time ,and , For the moment The observed state; Step 2: Input the flight trajectory historical observation sequence obtained in step 1 into the trained prediction model to predict the future flight trajectory sequence: in To predict the time length of the sequence, For the moment The predicted state of For the moment The predicted status.

6. The method according to claim 5, characterized in that: In step 1, ,in For at the moment When , the three-axis coordinate components of the flight trajectory point in the ground coordinate system are: For at the moment The three-axis velocity components of the flight trajectory point in the ground coordinate system when .

7. The method according to claim 5, characterized in that: The training process of the prediction model includes: Several segments of flight trajectory data are obtained as training sample data sets. For each training sample, it is divided into an observed trajectory and a future trajectory. The training sample is input into the prediction model, and the observed trajectory encoding network in the trajectory context feature encoding module is encoded. Parameters and differential position encoding network Parameters, parameters of the expert adapter layer in the pre-trained LLM module, and the network in the decoding module Parameters and Network Parameters for training.

8. The method according to claim 7, characterized in that: During the training process of the prediction model, the pre-trained LLM module adopts the uncertainty quantization mode: when each training sample is input for training, one expert adapter layer is randomly activated for training, and the other expert adapter layers are not activated; during actual prediction, the pre-trained LLM module adopts the precision mode: all expert adapter layers are activated.

9. The method according to claim 7, characterized in that: During the training process of the prediction model, the Huber loss function is used.

Images