Uncertain flight trajectory prediction model and method fused with flight dynamics knowledge

Abstract

The invention provides an uncertainty flight trajectory prediction model and method fusing flight dynamics knowledge. The prediction model comprises a trajectory context feature coding module, a pre-training LLM module and a decoding module based on a flight dynamics model. According to the method, the pre-trained large language model is introduced into the field of flight trajectory prediction for the first time, a special encoding-decoding process is designed for LLM by combining flight state characteristics and a flight dynamics model, the Flight LLM can give full play to the advantages of the LLM in time sequence data reasoning by means of the fusion, complex flight dynamics is modeled by means of the time sequence reasoning ability of the LLM, and the prediction efficiency of the LLM is improved. And meanwhile, the physical rationality of the trajectory is constrained through a kinetic model, so that complex dynamic characteristics in flight trajectory prediction are effectively represented. In addition, a multi-expert adapter is designed for the Flight LLM architecture, a trainable adapter is inserted into the pre-training LLM, and prediction precision improvement and uncertainty quantification are realized through integrated learning and random masks.

Description

Technical Field

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

[0002] Flight Trajectory Prediction (FTP) is a core technology of air traffic management (ATM). Its core goal is to accurately estimate the future flight status by analyzing the historical observation data of the aircraft (including spatiotemporal information such as position and speed). 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 types: physical model-driven and data-driven methods, each of which has significant technical characteristics and limitations:

[0003] The traditional physical model-driven FTP method is based on the flight dynamics equations and constructs a state evolution model by integrating aerodynamic / kinematic models with state estimation algorithms such as Kalman filtering. Some improved studies have improved prediction accuracy by introducing prior information such as flight intention recognition or estimated time of arrival (ETA). However, its effectiveness is limited by two essential defects: (1) the model mismatch phenomenon causes the prediction error to show a cumulative effect in multiple iterations; (2) the flight dynamics system has strong nonlinear characteristics and is coupled with a complex environment. The construction of a high-precision physical model faces engineering challenges such as difficulty in parameter identification and high computational complexity, which ultimately makes it difficult for the prediction accuracy to meet the actual application requirements.

[0004] With the development of machine learning technology, the data-driven FTP method based on multi-step multivariate time series prediction as a modeling framework has shown significant advantages. It directly extracts spatiotemporal 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 quantified uncertainty through Monte Carlo Dropout, Bayesian neural network and other methods. The current technological frontier presents three evolutionary directions: (1) Models based on the Transformer architecture use the self-attention mechanism to enhance the ability to capture long-range features; (2) Combine preprocessing techniques such as wavelet transform and trajectory binarization to optimize feature representation; (3) Attempt to embed flight dynamics constraints into the neural network structure design. Despite this, the existing methods still have room for improvement in terms of dynamic environment adaptability, physical interpretability and generalization ability of complex airspace scenarios.

[0005] Large language models (LLMs), represented by GPT / LLaMA, have shown unique potential in long sequence modeling and transfer learning due to their ultra-large-scale parameters and Transformer architecture advantages. Through domain adaptive training, LLM is expected to achieve dynamic correlation modeling of flight status information to improve FTP performance. However, practical applications face three challenges: (1) The matching degree between the highly dynamic characteristics of flight trajectories and the LLM temporal modeling capabilities is insufficiently verified; (2) The existing LLM framework lacks a reliable uncertainty quantification mechanism; (3) The fusion of prior knowledge of flight dynamics and data-driven models has not yet formed an effective architecture. Summary of the invention

[0006] In view of the technical problems existing in the existing data-driven FTP method, the present invention proposes an uncertain flight trajectory prediction model and method integrating flight dynamics knowledge, referred to as FlightLLM. This method combines the flight dynamics model with the 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 timing and semantic reasoning capabilities of LLM to effectively model complex flight dynamics.

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

[0008] An uncertain 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 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;

[0010] 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 ;

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

[0012] The network The high-dimensional feature sequence output by the pre-trained LLM module The last of Features Mapped to predicted overload instruction sequence ;

[0013] 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 ;

[0014] The network Coarse-grained flight trajectory sequence And high-dimensional feature sequences The last of Features is the input and output compensation term ;

[0015] The compensation module will compensate the Add 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 differential position encoding network ;

[0017] The observation trajectory encoding network The input flight trajectory history observation sequence Processing to generate high-dimensional feature vectors ;

[0018] 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;

[0019] 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 .

[0020] Furthermore, 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.

[0021] Furthermore, 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 .

[0022] The present invention also proposes an uncertain flight trajectory prediction method based on the above prediction model and integrating flight dynamics knowledge, comprising the following steps:

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

[0024]

[0025] in is the time length of the historical observation sequence, Indicates time ,and , For the moment The observed state;

[0026] 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:

[0027]

[0028] in is the time length of the predicted sequence, For the moment The predicted state of For the moment The predicted status.

[0029] Furthermore, 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 .

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

[0031] 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.

[0032] Furthermore, during the training of the prediction model, the pre-trained LLM module adopts an uncertainty quantization mode: when each training sample is input for training, one expert adapter layer is randomly activated for training, and the remaining expert adapter layers are not activated; during actual prediction, the pre-trained LLM module adopts a precision mode: all expert adapter layers are activated, and the knowledge of multiple expert adapter layer networks is aggregated through ensemble learning technology to optimize the prediction results.

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

[0034] Beneficial effects:

[0035] The present invention has the following effects:

[0036] 1. For the first time, a large language model-driven flight trajectory prediction method, FlightLLM, is proposed.

[0037] This invention proposes for the first time to introduce a pre-trained large language model into the field of flight trajectory prediction. Combining the flight state characteristics with the flight dynamics model, a dedicated encoding-decoding process is designed for LLM. This fusion enables FlightLLM to give full play to the advantages of LLM in time series data reasoning, and use the time series reasoning ability of LLM to model complex flight dynamics. At the same time, the physical rationality of the trajectory is constrained by the dynamics model, thereby effectively characterizing the complex dynamic characteristics in flight trajectory prediction.

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

[0039] This paper designs a multi-expert adapter for the FlightLLM architecture, inserts a trainable adapter into the pre-trained LLM, and achieves prediction accuracy improvement and uncertainty quantification through ensemble learning and random masking. The adapter can provide two reasoning modes:

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

[0041] (2) Uncertainty quantification model. This model provides more reliable predictions by evaluating prediction uncertainty and taking into account potential changes in flight trajectory. This model is particularly important in the aviation field, where uncertainty assessment is crucial for safety decision-making.

[0042] Experiments based on a dataset of real commercial flight trajectories around airports show that FlightLLM significantly outperforms existing models in prediction accuracy, while demonstrating effective uncertainty quantification capabilities, making it a potential tool for improving air traffic control safety and efficiency.

[0043] Additional aspects and advantages of the present invention will be given in part in the following description and in part will be obvious from the following description, or will be learned through 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 easily 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 : The overall architecture of FlightLLM;

[0047] Among them: (I) Trajectory context feature encoding module: encodes the observed flight trajectory into high-dimensional features that can be understood by LLM; (II) Pre-trained LLM module with multiple expert adapters: serves as the core architecture for deep temporal feature capture; (III) Decoding module based on flight dynamics model: decodes high-dimensional features into flight overload sequences through flight dynamics model to predict the state, and performs fine-grained compensation on the prediction results;

[0048] Figure 3 :Multi-Expert Adapter Architecture;

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

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

[0051] The lower right frame shows: In the integrated reasoning mode for improving prediction accuracy, multiple expert adapters are merged for joint computation;

[0052] Figure 4 : Schematic diagram of the predicted trajectory of each method in the xyz direction position component;

[0053] Among them: (a) low maneuvering ascending trajectory, (b) low maneuvering descending trajectory, (c) high maneuvering ascending trajectory, (d) high maneuvering descending trajectory;

[0054] Figure 5 : The predicted mean and 3σ upper and lower bounds of the sampling trajectory in the xyz direction based on multiple groups 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 under different prediction step sizes;

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

[0058] Figure 8 : Comparison of speed prediction performance results of various methods (15-step prediction);

[0059] Fig. 9 : Comparison of experimental results under different ablation scenarios (15-step prediction). DETAILED DESCRIPTION

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

[0061] (I) Implementation examples:

[0062] This embodiment is aimed at short-term FTP tasks, and its purpose is to predict the flight status in the future, such as position and speed, based on historical trajectory data, so as to improve the real-time safety management performance of high-density airspace such as airports. To this end, an uncertain 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: with pre-trained LLM as the backbone, it integrates prior knowledge of flight dynamics model and offline data, predicts future trajectory in an efficient non-autoregressive form, and supports uncertainty quantification. This method combines deep learning technology with domain knowledge to provide a robust and scalable solution for FTP.

[0063] The uncertain flight trajectory prediction method integrating 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 status through historical observation data. Therefore, we first obtain the historical observation sequence of the flight trajectory used for prediction:

[0065]

[0066] in is the time length of the historical observation sequence, Indicates time ,and , For the moment The observed state, ,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 ;

[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] in is the time length of the predicted sequence, For the moment The predicted state of For the moment The predicted state has the same meaning as the components in the observed state.

[0070] like Figure 2 As shown, the prediction model includes a trajectory context feature encoding module, a pre-trained LLM module containing multiple expert adapters, and a decoding module based on a 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). Figure 2 As shown in the first part, the trajectory context feature encoding module includes an observation trajectory encoding network and differential position encoding network ; The observation trajectory encoding network and the differential position encoding network They are composed of a GRU layer and an MLP layer respectively. The GRU layer captures the temporal features of the input sequence, and the MLP layer further fits it to generate a high-dimensional feature vector.

[0073] The flight trajectory history observation sequence Input 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 each moment , each predicted state in the coarse-grained flight trajectory sequence also consists of three-axis coordinate components and three-axis velocity components in the ground coordinate system.

[0080] like Figure 1 As 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. Therefore, by giving an initial state and combining the predicted overload instruction sequence, the coarse-grained future flight trajectory sequence can be recursively predicted through the three-degree-of-freedom particle model.

[0081] Because the three-degree-of-freedom particle model used is a simplified approximation of the actual flight process, its prediction has a large error. Therefore, the rough prediction results are further refined through a compensation mechanism. Specifically, the coarse-grained flight trajectory sequence High-dimensional feature sequence output by LLM module The last of Features Common Input Network , get compensation , the compensation item Add the coarse-grained flight trajectory sequence to get the final prediction result:

[0082]

[0083] Among them, the network It consists of two MLP layers and one GRU layer. and After each of the various fittings is passed through an MLP layer, the combined output is passed through the GRU layer to output the compensation term .

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

[0085] 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. And the differential position encoding network Parameters, parameters of the expert adapter layer in the pre-trained LLM module, and the network in the decoding module and network The parameters are trained, and 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 the uncertainty quantization mode, that is, 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.

[0087] In actual prediction, the pre-trained LLM module adopts the precision mode, activates all expert adapter layers, and aggregates the knowledge of multiple expert adapter layer networks through integrated learning technology to optimize the prediction results and significantly improve the accuracy.

[0088] (II) Example verification:

[0089] Example settings:

[0090] Dataset: In the example, we use the public flight trajectory dataset provided by the literature (Gariel, M.; Srivastava, AN; andFeron, 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 pre-processed data resampled at a 4-second interval, every 5 trajectory points are selected to create a trajectory segment with a length of 24 steps. Trajectories shorter than 24 steps are excluded, while trajectories of 24 steps or longer are divided into continuous, non-overlapping 24-step segments, and the remaining points are discarded. The final trajectory segment uses the first 9 steps as the model input and the last 15 steps as the predicted ground truth labels. 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 example verification, the comparative experiment selected two time series deep learning models commonly used in flight trajectory prediction (FTP): the autoregressive model and the parallel prediction model.

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

[0093] The parallel prediction model is based on the Transformer encoding-decoding structure and the LSTM architecture LSTM+Attention combined with the attention mechanism, and adopts a sequence-to-sequence (seq2seq) computing architecture.

[0094] Evaluation index setting: To evaluate the accuracy of position and speed estimation, this embodiment uses three core indicators:

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

[0096]

[0097] 2. Mean Absolute Error (MAE)

[0098]

[0099] 3. Mean Distance Error (MDE)

[0100]

[0101] Forecast Accuracy Analysis

[0102] In order to comprehensively evaluate the prediction accuracy of FlightLLM proposed in this example, the adapter is configured in ensemble inference mode and compared with other comparison algorithms at three different prediction lengths of 3, 9, and 15. For fairness, 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 method in dealing with long-term dependencies and trends not obvious. In the case of long-term prediction, the sequence-to-sequence method can better handle the long-term dependencies and complex dynamic characteristics of the trajectory by learning the long-term prediction process end-to-end. The complex maneuvers of the flying 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 the improvement of speed prediction performance, such as Figure 8 The velocity statistics of the 15-step prediction length are shown. Through reasoning based on the prior dynamics model, FlightLLM effectively captures the coupling relationship between position and velocity, thus performing outstandingly in velocity prediction.

[0104] Figure 7The prediction performance of each method in the take-off and landing phases is further demonstrated. For the flight trajectory prediction problem near the airport, the prediction difficulty of the take-off trajectory is significantly higher than that of the landing trajectory. During the landing phase, the aircraft usually has a clear destination, and its movement is more geographically sensitive and easier to learn from offline data. However, the take-off process of the aircraft has more diverse changes, resulting in lower overall prediction accuracy. Obviously, FlightLLM's powerful data learning ability enables it to capture the motion characteristics of the target in different flight phases, thereby achieving accurate predictions for both take-off and landing phases, and the prediction accuracy is better than other methods.

[0105] In addition, Figure 4 In the figure, four groups of sub-figures are used to show the comparison of trajectory prediction results of FlightLLM with other methods (such as Bi-LSTM, Transformer, etc.) in different maneuvering scenarios:

[0106] Low maneuvering scenario (a, b): The aircraft moves smoothly and the trajectory changes little. The predicted trajectory (dashed line) of FlightLLM (red) is highly consistent with the true trajectory (black solid line), while other methods (such as blue and green) gradually deviate in long-term prediction;

[0107] High maneuverability scenarios (c, d): The aircraft moves in a complex manner and the trajectory changes dramatically. FlightLLM significantly suppresses the error accumulation of traditional methods (such as Bi-GRU) by combining dynamic priors; 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 and high maneuverability situations. In contrast, the autoregressive BLSTM and BGRU show significant prediction deviations due to error accumulation. Although LSTM+Attention and Transformer can better learn the laws of flight motion changes, their overall prediction performance is still inferior to FlightLLM. More importantly, by learning overload instructions and predicting trajectories based on dynamic models, FlightLLM's predicted trajectory is constrained by flight dynamics, which improves accuracy while ensuring the physical rationality of the predicted trajectory. Figure 4 The advantages of FlightLLM in complex maneuver adaptability and long-term prediction stability were verified, and its physical constraint mechanism (dynamic model) effectively improved the rationality of the trajectory.

[0109] Prediction uncertainty analysis

[0110] In order to evaluate the ability of FlightLLM to quantify the uncertainty of prediction, 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 calculated respectively. Figure 5The 3D visualization shows the sampling trajectory and describes the predicted mean and ±3σ boundaries in the xyz direction (σ is the standard deviation). Obviously, the initial prediction accuracy is high, and the sampling trajectory predicted by FlightLLM has a small dispersion, indicating that its prediction uncertainty is low. Figure 5 The sampled trajectories in the early and middle stages almost overlap, and the corresponding xyz direction discreteness is small. As the prediction time increases, the aircraft's climb is usually accompanied by a change in motion pattern, resulting in a decrease in the accuracy of the later predictions. At this time, the sampled trajectories of FlightLLM have higher discreteness and larger variance, enabling it to output diversified prediction results. Therefore, FlightLLM can effectively measure prediction uncertainty, evaluate prediction risks, and identify abnormal motion patterns in trajectories through changes in discreteness or variance.

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

[0112] Sampling trajectories: Generate multiple possible trajectories by randomly activating different expert adapters (multi-expert structure).

[0113] Confidence interval: In the initial stage (first 5 steps), the forecast uncertainty (σ value) is small and the confidence interval is narrow, indicating that the model has high confidence in the short-term forecast; as the forecast step increases (latter 10 steps), the confidence interval gradually expands, reflecting the natural growth of uncertainty.

[0114] Key findings:

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

[0116] The true trajectory (black) always lies within the 3σ interval, demonstrating the reliability of the uncertainty estimate.

[0117] Effectiveness analysis of each module

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

[0119] 1. Remove the multi-expert adapters in FlightLLM, so that LLM can be used as a single backbone network without fine-tuning through adapters;

[0120] 2. Exclude the dynamic model output and directly Input a single-layer MLP with 128 nodes to output the predicted trajectory;

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

[0122] like Fig. 9 As shown in the figure, the prediction performance of these three ablation scenarios is tested at a prediction length of 15 steps. Obviously, the absence of any module will lead to a decrease in the prediction performance of FlightLLM. First, the absence of the multi-expert adapter has the greatest impact on the overall performance, because it enhances the model's ability to understand and learn complex features through fine-tuning, thereby improving its adaptability to complex tasks and actual data. The absence of the dynamic model significantly affects the speed prediction performance because it effectively captures and utilizes the coupling relationship between speed and position. In addition, differential query also has a significant impact because it introduces dynamic change information at different time scales in addition to 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, which uses a pre-trained large language model (LLM) as the core architecture and achieves efficient short-term flight trajectory prediction by integrating prior knowledge of flight dynamics models. We designed a new multi-expert adapter that supports prediction uncertainty quantification while improving prediction accuracy. A large number of experiments have shown that FlightLLM can achieve excellent prediction accuracy under different maneuvering intensities and effectively quantify prediction uncertainty, thereby ensuring that trajectory prediction has both high accuracy and high reliability. Overall, FlightLLM provides a promising technical tool for improving air traffic control safety and operational efficiency by providing accurate and reliable trajectory predictions.

[0124] Although the embodiments of the present invention have been shown and described above, it is to be understood that the above embodiments are exemplary and are not to be construed as limitations on the present invention. A person skilled in the art may change, modify, replace and modify the above embodiments within the scope of the present invention without departing from the principles and intent 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 by: 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 GRU layer captures the temporal features of the input sequence and is 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