Real-time aviation flow prediction method based on aviation situation awareness

Abstract

The invention discloses a real-time aviation flow prediction method based on aviation situation awareness, and the method comprises the steps: constructing a flow data set fusing spatial-temporal characteristics, and achieving the efficient conversion from trajectory data to aviation flow data based on time slice division; extracting multi-source spatio-temporal features; a light GBM is adopted to construct a prediction model, and the core lies in that a Tweedie loss function is used to model the zero expansion and long tail distribution characteristics of flow data, and the robustness of the model is ensured through five-fold cross validation; establishing a multi-dimensional evaluation system; in the application stage, the airspace is divided into dynamic grids, and flight traffic in a plurality of minutes in the future in the airspace is predicted in real time in combination with radar return trajectory data. The method supports dynamic feature extraction and flow prediction based on real-time radar data, can effectively support an aviation flow management and control decision through actual measurement verification, can quickly respond when the airspace flow changes, meets the timeliness demand of dynamic management and control of air traffic control, and remarkably improves the scientificity and response speed of air traffic control decision.

Description

Technical Field

[0001] The present invention relates to the technical field of air traffic management, and in particular to a real-time aviation traffic flow prediction method based on aviation situation awareness, which belongs to the cross-application technology of intelligent transportation, aviation data mining and traffic flow management. Background Art

[0002] With the rapid development of the air transport industry and the continuous growth of flight traffic within airspace, accurate forecasting of air traffic flow has become a critical component in ensuring flight safety, improving airspace utilization, and optimizing air traffic control resource allocation. Existing air traffic flow forecasting methods primarily include traditional methods based on statistical analysis and modern methods based on machine learning. Traditional methods, such as time series analysis (ARIMA), have limitations in handling complex spatiotemporal characteristics and nonlinear relationships, making them difficult to adapt to the sudden and volatile nature of flight traffic. While some machine learning-based models have improved forecasting accuracy, they still lack advantages in feature integration, model generalization, and adaptability to zero-inflation and long-tail distributions of traffic data. Furthermore, existing methods often rely on a single metric in their model evaluation systems, lacking comprehensive consideration of multiple dimensions such as peak / off-peak periods and error distribution. This results in limited interpretability and practical application in actual air traffic control decision-making. Therefore, there is an urgent need for an air traffic flow forecasting method that fully integrates spatiotemporal characteristics, possesses efficient learning capabilities, adapts to the distribution characteristics of traffic data, and provides a comprehensive evaluation system. Summary of the Invention

[0003] The purpose of the present invention is to provide a real-time aviation traffic prediction method based on aviation situation awareness to overcome the shortcomings of the existing technology.

[0004] A real-time aviation traffic flow prediction method based on aviation situation awareness is characterized by comprising the following steps: (1) Constructing spatiotemporal fusion dataset: a) Using radar track reports sent by aircraft during flight as the original radar track dataset ,in Is a set of flight identifiers, each flight trajectory ,in Indicates that the trajectory contains sampling points, is the timestamp, is the three-dimensional space coordinate, abbreviated as ; b) with fixed time slices (like 1 minute) to slice the time of one day and obtain a set of time slices ; c) In each time slice Within, through statistical airspace The number of flights in a time slice is used to represent the airspace traffic in that time slice; (2) Extraction of multidimensional features: a) Time characteristics: including hour code and day of the week code , Holiday Code ; b) Seasonal characteristics: month and quarter codes; c) Historical traffic characteristics: in the past Traffic sequence of time slices ; d) Traffic statistics: based on past time windows Calculate the mean and variance, , , e) Traffic trend characteristics: the slope of the historical traffic series, , (3) Construction of traffic prediction model based on LightGBM: a) Using the features extracted in step (2) as input and the spatial traffic volume as labels, a supervised learning dataset is constructed; The LightGBM (Light Gradient Boosting Machine) framework is used for model training and verification. The model input is passed in step (2) c) The flow sequence of each time slice is the main feature, and the time, season, flow statistics and flow trend features of step (2) a), b), d) and e) are auxiliary features. The model output is the future Predicted airspace traffic within minutes , and It is a hyperparameter, determined during model training; b) The loss function uses the Tweedie loss function, which is defined as: , in For the n The actual traffic of the samples, For the n The predicted flow of samples, is the total number of samples, is the Tweedie distribution parameter, ,This loss function can effectively handle the zero-inflated and long-tail distribution characteristics of aviation traffic data; d) Five-fold cross-validation (K=5) was used for model training and evaluation to ensure model robustness and generalization ability; (4) Real-time prediction of airspace traffic in the application stage: Airspace , based on real-time radar trajectory data and trained LightGBM model, predict the future Airspace traffic after minutes: , in minute, is the number of consecutive time slices in the past, Indicates airspace Internal Past Continuous The features extracted by step (2) correspond to the time slices.

[0005] In step (1), step c) counts the airspace flow in the following way, and defines the airspace flow as: , in is the indicator function, defined as: .

[0006] After step c) in step (1), there is step d) airspace Dynamic division of airspace Select several sub-spaces , let the grid set be ; Change the step c) in step (1) to Replace with The sub-space domain can be calculated Airspace traffic within.

[0007] The step (4) is to perform a , based on real-time radar trajectory data and trained LightGBM model, predict the future Airspace traffic within minutes: , in minute, is the number of consecutive time slices in the past; Represents a sub-airspace Internal Past Continuous The features extracted by step (2) correspond to each time slice.

[0008] The present invention achieves a flow-based expression of flight trajectories by downsampling and time-slicing the raw radar trajectory data, combined with spatial grid division. It also integrates multidimensional features such as time, historical traffic, statistics, trends, and seasons into feature engineering, significantly improving the model's ability to model complex spatiotemporal relationships. LightGBM is used as the core prediction model, leveraging its efficient learning mechanism based on gradient boosting decision trees to automatically screen important features and adapt to large-scale, high-dimensional data. By introducing the Tweedie loss function, the model can effectively handle the zero-inflation and long-tail distribution problems commonly found in traffic data, improving the accuracy and robustness of predictions.

[0009] This paper proposes a multi-dimensional evaluation standard, including mean absolute error (MAE), error quantiles, peak / off-peak error comparison, and feature importance ranking. This system not only comprehensively reflects the overall performance of the model but also reveals its key performance in actual air traffic control applications, facilitating targeted optimization and decision support, and improving the model's interpretability and practical value. This paper supports dynamic feature extraction and traffic flow prediction based on real-time radar data, enabling rapid response to changes in airspace traffic flow, meeting the timeliness requirements of dynamic air traffic control, and significantly improving the scientific nature and responsiveness of air traffic control decisions. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] Figure 1 Overall flow chart of the present invention.

[0011] Figure 2 A visualization of the prediction error comparison.

[0012] Figure 3 Schematic diagram of the prediction error distribution visualization.

[0013] Figure 4 Schematic diagram of feature importance visualization.

[0014] Figure 5 Schematic diagram comparing the original sequence of actual values ​​and predicted values.

[0015] Figure 6 Schematic diagram of MAE error visualization for each time period. DETAILED DESCRIPTION

[0016] like Figure 1 As shown, a real-time aviation traffic prediction method based on aviation situational awareness is characterized by comprising the following steps: (1) Constructing spatiotemporal fusion dataset: a) Using radar track reports sent by aircraft during flight as the original radar track dataset ,in Is a set of flight identifiers, each flight trajectory , which means that the trajectory contains sampling points, is the timestamp, is the three-dimensional space coordinate, abbreviated as .

[0017] b) with fixed time slices (For example, 1 minute) Slice the time of a day to obtain a set of time slices .

[0018] c) In each time slice Within, through statistical airspace The number of flights in a time slice is used to represent the airspace flow in the time slice; the airspace flow is defined as: , in is the indicator function, defined as: .

[0019] d) Airspace Dynamic division of airspace Select several sub-spaces , let the grid set be ; Change the step c) in step (1) to Replace with The sub-space domain can be calculated Airspace traffic within.

[0020] (2) Extraction of multidimensional features: a) Time characteristics: including hour code , week code , Holiday Code ; b) Seasonal characteristics: month and quarter codes; c) Historical traffic characteristics: in the past Traffic sequence of time slices ; d) Traffic statistics: based on past time windows Calculate the mean and variance, , , e) Traffic trend characteristics: the slope of the historical traffic series, .

[0021] (3) Construction of traffic prediction model based on LightGBM: a) Using the features extracted in step (2) as input and the spatial traffic volume as labels, a supervised learning dataset is constructed; The LightGBM (Light Gradient Boosting Machine) framework is used for model training and verification. The model input is passed in step (2) c) The flow sequence of each time slice is the main feature, and the time, season, flow statistics and flow trend features of step (2) a), b), d) and e) are auxiliary features. The model output is the future Predicted airspace traffic within minutes , and It is a hyperparameter, determined during model training; b) The loss function uses the Tweedie loss function, which is defined as: , in For the n The actual traffic of the samples, For the n The predicted flow of samples, is the total number of samples, is the Tweedie distribution parameter, ,This loss function can effectively handle the zero-inflated and long-tail distribution characteristics of aviation traffic data; d) Five-fold cross-validation (K=5) is used for model training and evaluation to ensure model robustness and generalization ability.

[0022] (4) Establishment of a multi-dimensional evaluation system: a) Build a comprehensive model evaluation standard, using mean absolute error (MAE) as the main evaluation metric, defined as: , b) Perform error quantile analysis to calculate the quantiles of the prediction error (e.g., 50% and 90% quantiles) to reflect the performance of the model at different error levels; c) Peak / off-peak period error comparison, calculate the peak period (such as hourly flow greater than the threshold and MAE during off-peak hours to ensure the model’s prediction reliability during critical periods; d) Feature importance ranking, based on the feature importance output of LightGBM, ranking the contribution of features to the prediction; e) The above multi-dimensional evaluation criteria can not only comprehensively reflect the overall performance of the model, but also reveal the key performance of the model in actual air traffic control applications, facilitate targeted optimization and decision support, and enhance the interpretability and practical value of the model.

[0023] (5) Real-time prediction of airspace traffic in the application phase: a) Airspace , based on real-time radar trajectory data and trained LightGBM model, predict the future Airspace traffic after minutes: , in minute, Past continuous A time slice, Indicates airspace Internal Past Continuous The features extracted by step (2) correspond to each time slice.

[0024] b) When the airspace is dynamically divided, each sub-airspace , based on real-time radar trajectory data and trained LightGBM model, predict the future Airspace traffic within minutes: , in minute, Past continuous time slices; Represents a sub-airspace Internal Past Continuous The features extracted by step (2) correspond to each time slice.

[0025] This method has the characteristics of real-time prediction. After receiving the latest radar data, it can quickly complete feature extraction and traffic prediction, meet the timeliness requirements of dynamic control of airspace traffic, and significantly improve the scientific nature and response speed of air traffic control decisions.

[0026] Example: In this embodiment, the radar return track data of the controlled flights of the approach department / area department of a domestic airport from January to April 2025 is used. According to statistics, there are more than 200,000 tracks in total. Each track has several track points, and the track point time interval is 4 seconds. It consists of several tags returned by the radar during the flight of the aircraft. For ease of understanding, the present invention simplifies the tags used into mathematical logic. Each track contains the flight index f , timestamp t ,longitude x ,latitude y ,high z This embodiment mainly includes the following 5 steps, see the overall flow chart Figure 1 .

[0027] (1) Construction of spatiotemporal fusion dataset: a) The radar track reports sent by the aircraft during flight are called the original radar track data set. Since the original data set is track data, it needs to be converted into an airspace traffic data set. The sampling interval for the original data set is 4 seconds. If it is directly converted into a traffic data set, a large amount of duplicate data will be generated, resulting in information redundancy, which is not conducive to model training. Therefore, the present invention first downsamples the original radar track data on a real time scale, with a downsampling interval of 1 minute. Assume that the original radar track data set ,in Is a set of flight identifiers, for flight trajectories Include sampling points, is the timestamp, is the three-dimensional space coordinate, abbreviated as .

[0028] b) Now set a fixed time slice for each downsampling time point = 8s to slice the trajectory data and obtain a set of time slices .

[0029] c) In each time slice Within, through statistical airspace The number of flights in a time slice is used to represent the airspace traffic within the time slice. For example, the airspace at 01:01 on January 1st is counted. Airspace traffic within, select all radar track reports between 01:01:00 and 01:01:08 and determine their longitude x ,latitude y ,high z Is the altitude in the airspace? If in the airspace Internal airspace traffic +1, each flight index f It can only be recorded once in each time slice.

[0030] d) Airspace Divided into multiple airspace sectors, the sector set is .

[0031] (2) Extraction of multidimensional features: In order to improve the model's ability to model complex spatiotemporal relationships, multi-dimensional feature extraction is performed on the constructed traffic data set. , the extracted features include: a) Time characteristics: including hour code , week code , Holiday Code ,The time feature is encoded using one-hot encoding; b) Seasonal characteristics: Month and quarter codes, with monthly and seasonal characteristics using one-hot encoding; c) Historical traffic characteristics: in the past Traffic sequence of time slices Recorded as ; d) Traffic statistics: Calculate the mean and variance based on the time window of the past 15 consecutive time slices. , , e) Traffic trend characteristics: the slope of the historical traffic series, .

[0032] (3) Construction of traffic prediction model based on LightGBM: a) Features extracted in step (2) As input, airspace traffic As labels, construct supervised learning datasets.

[0033] b) Use LightGBM (Light Gradient Boosting Machine) framework for model training and verification. The model input is based on the past The flow sequence of each time slice is the main feature, and time, season, flow statistics and flow trend features are auxiliary features. The predicted traffic volume for the airspace within one minute.

[0034] c) The loss function uses the Tweedie loss function. In this example, This loss function can effectively handle the zero-inflation and long-tail distribution characteristics of aviation traffic data; d) Five-fold cross-validation (K=5) is used for model training and evaluation to ensure model robustness and generalization ability.

[0035] (4) Establishment of a multi-dimensional evaluation system: a) Build a comprehensive model evaluation standard, using mean absolute error (MAE) as the main evaluation metric; b) Perform error quantile analysis to calculate the quantiles of the prediction error (e.g., 50% and 90% quantiles) to reflect the performance of the model at different error levels; c) Peak / off-peak period error comparison, calculate the peak period (such as hourly flow greater than the threshold and MAE during off-peak hours to ensure the model’s prediction reliability during critical periods; d) Feature importance ranking, based on the feature importance output of LightGBM, ranking the contribution of features to the prediction; e) The above multi-dimensional evaluation criteria can not only comprehensively reflect the overall performance of the model, but also reveal the key performance of the model in actual air traffic control applications, facilitate targeted optimization and decision support, and improve the interpretability and practical value of the model. Figure 2-Figure 6 . Comparison between predicted and actual values ​​( Figure 2 ) point of view, most of the points are distributed near the fitting line, indicating that the overall prediction effect of the model is good and can more accurately reflect the actual flow change trend; from the perspective of prediction error distribution ( Figure 3 ) perspective, the error distribution is approximately symmetrical, with the center close to zero, indicating that the model as a whole does not have systematic bias, the prediction errors are mainly concentrated in a small range, and there are fewer extreme errors, which further verifies the stability and reliability of the model; from the perspective of information gain ( Figure 4 ) perspective, it can be seen that historical traffic features (such as past_inside_flow and max) contribute the most to model predictions, while time features (such as hour, minute, and dayofyear) and statistical features (such as mean and median) are also highly important. This shows that the model can effectively utilize multi-dimensional spatiotemporal features to improve prediction accuracy. Finally, based on the comparison chart of the original sequence of actual values ​​and predicted values ​​( Figure 5 ), it can be seen that the model can better capture the temporal variation of traffic, and the prediction results are highly consistent with the actual observations.

[0036] During the evaluation phase, we deliberately deleted most of the evaluation samples during the early morning hours when the traffic was 0, because these hours are highly predictable and the basic traffic is 0. If we retain these evaluation samples with 0 values, the overall MAE will be very good (the overall MAE is about 0.55), but this is very misleading in actual applications. Figure 6 The five-fold cross-validation curve shows that the largest prediction errors occur between 10:00 and 11:00 and 18:00 and 19:00, which are also the periods of the day with the largest flight traffic. This is consistent with empirical intuition. The MAE errors during these periods are approximately between 2 and 3.

[0037] therefore, 1. If a large number of evaluation samples with 0 values ​​are retained, the overall MAE will be particularly good (about 0.55), but it cannot reflect the actual situation; 2. Removing a large number of evaluation samples with zero values ​​will result in a less attractive overall MAE (approximately 2.5), but it better reflects the performance of actual flights during regular operations. This analysis helps identify differences in model performance during peak and off-peak hours, providing data support for subsequent model optimization and actual air traffic control applications.

[0038] These experimental results fully demonstrate the effectiveness and practicality of the proposed method for real-time traffic flow prediction based on aviation trajectories. The model not only achieves high overall prediction accuracy and a reasonable error distribution, but also identifies key features and adapts to traffic flow variations over time, demonstrating promising practical application prospects.

[0039] (5) Real-time prediction of airspace traffic in the application phase: a) In step (1), the airspace is divided into dynamic grids. Let the grid set be ; b) For each airspace sector , based on real-time radar trajectory data and the trained LightGBM model (step (3)), predict the future Airspace traffic within minutes: , in minute, Past continuous time slices; c) This method offers real-time prediction capabilities, enabling rapid feature extraction and flow prediction upon receiving the latest radar data. This approach meets the timeliness requirements of dynamic airspace flow control and significantly improves the scientific nature and responsiveness of air traffic control decisions. Our invention, combined with the city's Kafka real-time trajectory flow system, has been deployed at this domestic airport.

Claims

1. A real-time aviation traffic prediction method based on aviation situational awareness, characterized by The steps include: (1) Constructing spatiotemporal fusion dataset: a) Using radar track reports sent by aircraft during flight as the original radar track dataset , where is a set of flight identifiers, each flight trajectory ,in Indicates that the trajectory contains sampling points, is the timestamp, is the three-dimensional space coordinate, abbreviated as ; b) with fixed time slices Slice the time of a day to get a set of time slices ; c) In each time slice Within, through statistical airspace The number of flights in a time slice is used to represent the airspace traffic in that time slice; (2) Extraction of multidimensional features: a) Time characteristics: including hour code , week code , Holiday Code ; b) Seasonal characteristics: month and quarter codes; c) Historical traffic characteristics: in the past Traffic sequence of time slices ; d) Traffic statistics: based on past time windows Calculate the mean and variance, ， ， e) Traffic trend characteristics: the slope of the historical traffic series, ， (3) Construction of traffic prediction model based on LightGBM: a) Using the features extracted in step (2) as input and the spatial traffic volume as labels, a supervised learning dataset is constructed; The LightGBM (Light Gradient Boosting Machine) framework is used for model training and verification. The model input is passed in step (2) c) The flow sequence of each time slice is the main feature, and the time, season, flow statistics and flow trend features of step (2) a), b), d) and e) are auxiliary features. The model output is the future Predicted airspace traffic within minutes , and It is a hyperparameter, determined during model training; b) The loss function uses the Tweedie loss function, which is defined as: ， in For the n The actual traffic of the samples, For the n The predicted flow of samples, is the total number of samples, is the Tweedie distribution parameter, ,This loss function can effectively handle the zero-inflated and long-tail distribution characteristics of aviation traffic data; d) Five-fold cross-validation (K=5) was used for model training and evaluation to ensure model robustness and generalization ability; (4) Real-time prediction of airspace traffic in the application stage: Airspace , based on real-time radar trajectory data and trained LightGBM model, predict the future Airspace traffic after minutes: ， in minute, is the number of consecutive time slices in the past, Indicates airspace Internal Past Continuous The features extracted by step (2) correspond to each time slice.

2. The method according to claim 1, wherein The fixed time slice described in step (1) For 1 minute.

3. The method according to claim 1, wherein In step (1), step c) counts the airspace flow in the following way, and defines the airspace flow as: ， in is the indicator function, defined as: 。 4. The method according to claim 1 or 3, wherein After step c) in step (1), there is step d) airspace Dynamic division of airspace Select several sub-spaces , let the grid set be ; Change the step c) in step (1) to Replace with The sub-space domain can be calculated Airspace traffic within.

5. The method according to claim 4, wherein Step (4) is to perform , based on real-time radar trajectory data and trained LightGBM model, predict the future Airspace traffic within minutes: ， in minute, is the number of consecutive time slices in the past; Represents a sub-airspace Internal Past Continuous The features extracted by step (2) correspond to each time slice.

Images