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

Abstract

A real-time air traffic flow prediction method based on aviation situational awareness includes: constructing a traffic flow dataset that integrates spatiotemporal features; achieving efficient conversion from trajectory data to air traffic flow data based on time slice partitioning; extracting multi-source spatiotemporal features; constructing a prediction model using LightGBM, the core of which lies in: using the Tweedie loss function to model the zero-inflation and long-tail distribution characteristics of traffic flow data, and ensuring model robustness through five-fold cross-validation; establishing a multi-dimensional evaluation system; and in the application phase, dividing the airspace into a dynamic grid and combining radar-returned trajectory data to predict flight traffic flow within the airspace for the next few minutes in real time. This invention supports dynamic feature extraction and traffic flow prediction based on real-time radar data. Practical testing has verified that it effectively supports air traffic flow control decisions, enabling rapid response when airspace traffic changes, meeting the timeliness requirements of dynamic air traffic control, and significantly improving the scientific nature and response speed of air traffic control decisions.

Description

Technical Field

[0001] This invention relates to the field of air traffic management technology, specifically to a real-time air traffic flow prediction method based on aviation situational awareness, which belongs to the cross-application technology of intelligent transportation, aviation data mining and traffic management. Background Technology

[0002] With the rapid development of the air transport industry, air traffic flow has continued to grow, making accurate prediction of air traffic flow a crucial link in ensuring flight safety, improving airspace utilization, and optimizing air traffic control resource allocation. Existing air traffic flow prediction methods mainly include traditional methods based on statistical analysis and modern methods based on machine learning. Traditional methods, such as time series analysis (ARIMA, etc.), have limitations in handling complex spatiotemporal characteristics and nonlinear relationships, making them difficult to adapt to the suddenness and variability of air traffic flow. While some machine learning-based models have improved prediction accuracy, they still have shortcomings in feature fusion, model generalization ability, and adaptability to zero-inflation and long-tail distribution of traffic data. Furthermore, existing methods often rely on single indicators in their model evaluation systems, lacking comprehensive consideration of peak / off-peak periods, error distribution, and other dimensions, resulting in limited interpretability and practicality of the models in actual air traffic control decision-making. Therefore, there is an urgent need for an air traffic flow prediction method that can fully integrate spatiotemporal characteristics, possess efficient learning capabilities, adapt to the distribution characteristics of traffic data, and have a comprehensive evaluation system. Summary of the Invention

[0003] The purpose of this invention is to provide a real-time air traffic forecasting method based on aviation situational awareness, so as to overcome the shortcomings of the prior art.

[0004] A real-time air traffic forecasting method based on aviation situational awareness, characterized by the following steps:

[0005] (1) Constructing a spatiotemporal fusion dataset:

[0006] a) Use radar trajectory reports sent by the aircraft during flight as the original radar trajectory dataset. ,in It is a set of flight identifiers, and each flight trajectory ,in This indicates that the trajectory contains One sampling point, For timestamps, For three-dimensional spatial coordinates, abbreviated as ;

[0007] b) Using fixed time slices (like Divide a day into time slices (each slice is 1 minute long) to obtain a set of time slices. ;

[0008] c) In each time slice Within, through statistical airspace The number of flights within a given time frame represents the airspace traffic during that time segment.

[0009] (2) Extraction of multidimensional features:

[0010] a) Time characteristics: including hour code and weekday code Holiday codes ;

[0011] b) Seasonal characteristics: Month and quarter codes;

[0012] c) Historical flow characteristics: past Flow sequence of time slices ;

[0013] d) Traffic statistics characteristics: based on past time windows Calculate the mean and variance.

[0014] ,

[0015] ,

[0016] e) Flow trend characteristics: the slope of historical flow sequences,

[0017] ,

[0018] (3) Construction of a traffic prediction model based on LightGBM:

[0019] a) Using the features extracted in step (2) as input and spatial traffic as labels, construct a supervised learning dataset;

[0020] The LightGBM (Light Gradient Boosting Machine) framework was used for model training and validation. The model input was from step (2) c). The model output is the flow sequence of each time slice as the main feature, and the time, season, flow statistics and flow trend features of steps (a), (b), (d) and (e) of step (2) as auxiliary features. Predicted airspace flow within minutes , and These are hyperparameters, determined during model training;

[0021] b) The Tweedie loss function is used as the loss function, defined as:

[0022] ,

[0023] in For the n The actual traffic of each sample For the n Predicted flow for each sample The total number of samples, For Tweedie distribution parameters, This loss function can effectively handle the zero-inflation and long-tail distribution characteristics of air traffic data;

[0024] d) Five-fold cross-validation (K=5) is used for model training and evaluation to ensure the model's robustness and generalization ability;

[0025] (4) Real-time airspace traffic prediction during the application phase:

[0026] airspace Based on real-time radar trajectory data and a trained LightGBM model, predict the future. Airspace traffic minutes later:

[0027] ,

[0028] in minute, The number of consecutive time slices in the past. Represents airspace Inner past continuous The features extracted in step (2) correspond to each time slice.

[0029] In step (1), step c) uses the following method to calculate airspace traffic, defining airspace traffic as:

[0030] ,

[0031] in The indicator function is defined as follows:

[0032] .

[0033] Step (1) c) is followed by step d) airspace. Dynamic partitioning, from airspace Select several sub-space domains Let the grid set be ; in step (1) c) Replace with Then the subspace can be calculated. Airspace traffic within the region.

[0034] Step (4) involves processing each subspace. Based on real-time radar trajectory data and a trained LightGBM model, predict the future. Airspace traffic within minutes:

[0035] ,

[0036] in minute, The number of consecutive time slices in the past; Represents subspace Inner past continuous The features extracted in step (2) correspond to each time slice.

[0037] This invention achieves a flow-oriented representation of flight trajectories by downsampling and time-slicing the original radar trajectory data, combined with spatial grid partitioning. Furthermore, it integrates multi-dimensional features such as time, historical flow, statistics, trends, and seasons during feature engineering, significantly improving the model's ability to model complex spatiotemporal relationships. Employing LightGBM as the core prediction model, its efficient learning mechanism based on gradient boosting decision trees automatically selects important features, adapting to large-scale, high-dimensional data. By introducing the Tweedie loss function, the model effectively handles the common zero-inflation and long-tail distribution problems in flow data, improving prediction accuracy and robustness.

[0038] This invention proposes a multi-dimensional evaluation standard including mean absolute error (MAE), error quantiles, peak / off-peak period error comparison, and feature importance ranking. This system not only comprehensively reflects the overall model performance but also reveals key performance characteristics in practical air traffic control applications, facilitating targeted optimization and decision support, and enhancing the model's interpretability and practical value. This invention supports dynamic feature extraction and traffic forecasting based on real-time radar data, enabling rapid response to changes in airspace traffic, meeting the timeliness requirements of dynamic air traffic control, and significantly improving the scientific rigor and response speed of air traffic control decisions. Attached Figure Description

[0039] Figure 1 The overall flowchart of this invention.

[0040] Figure 2 A visual illustration comparing prediction errors.

[0041] Figure 3 A visual representation of the prediction error distribution.

[0042] Figure 4 A visualization of feature importance.

[0043] Figure 5A diagram showing the comparison between the original sequences of actual and predicted values.

[0044] Figure 6 Visualization diagram of MAE error at different time periods. Detailed Implementation

[0045] like Figure 1 As shown, a real-time air traffic forecasting method based on aviation situational awareness is characterized by the following steps:

[0046] (1) Constructing a spatiotemporal fusion dataset:

[0047] a) Use radar trajectory reports sent by the aircraft during flight as the original radar trajectory dataset. ,in It is a set of flight identifiers, and each flight trajectory , where indicates that the trajectory contains One sampling point, For timestamps, For three-dimensional spatial coordinates, abbreviated as .

[0048] b) Using fixed time slices (For example, 1 minute) Divide a day into time slices to obtain a set of time slices. .

[0049] c) In each time slice Within, through statistical airspace The airspace traffic within a given time slice is represented by the number of flights within that slice; airspace traffic is defined as:

[0050] ,

[0051] in The indicator function is defined as follows:

[0052] .

[0053] d) Airspace Dynamic partitioning, from airspace Select several sub-space domains Let the grid set be ; in step (1) c) Replace with Then the subspace can be calculated. Airspace traffic within the region.

[0054] (2) Extraction of multidimensional features:

[0055] a) Time characteristics: including hourly coding Weekday Code Holiday codes ;

[0056] b) Seasonal characteristics: Month and quarter codes;

[0057] c) Historical flow characteristics: past Flow sequence of time slices ;

[0058] d) Traffic statistics characteristics: based on past time windows Calculate the mean and variance.

[0059] ,

[0060] ,

[0061] e) Flow trend characteristics: the slope of historical flow sequences,

[0062] .

[0063] (3) Construction of a traffic prediction model based on LightGBM:

[0064] a) Using the features extracted in step (2) as input and spatial traffic as labels, construct a supervised learning dataset;

[0065] The LightGBM (Light Gradient Boosting Machine) framework was used for model training and validation. The model input was from step (2) c). The model output is the flow sequence of each time slice as the main feature, and the time, season, flow statistics and flow trend features of steps (a), (b), (d) and (e) of step (2) as auxiliary features. Predicted airspace flow within minutes , and These are hyperparameters, determined during model training;

[0066] b) The Tweedie loss function is used as the loss function, defined as:

[0067] ,

[0068] in For the n The actual traffic of each sample For the n Predicted flow for each sample The total number of samples, For Tweedie distribution parameters, This loss function can effectively handle the zero-inflation and long-tail distribution characteristics of air traffic data;

[0069] d) Five-fold cross-validation (K=5) is used for model training and evaluation to ensure the robustness and generalization ability of the model.

[0070] (4) Establishment of a multi-dimensional evaluation system:

[0071] a) Construct a comprehensive model evaluation standard, using mean absolute error (MAE) as the main evaluation indicator, defined as:

[0072] ,

[0073] b) Perform error quantile analysis to calculate the quantiles of the prediction error (e.g., 50% and 90% quantiles) to reflect the model's performance at different error levels;

[0074] c) Comparison of peak / off-peak period errors, calculating the peak period (e.g., hourly flow rate exceeding the threshold) separately. MAE during off-peak hours ensures the reliability of model predictions during critical periods;

[0075] d) Feature importance ranking: Based on the feature importance output of LightGBM, the contribution of ranked features to the prediction is determined.

[0076] 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, which facilitates targeted optimization and decision support, and improves the interpretability and practical value of the model.

[0077] (5) Real-time airspace traffic prediction during the application phase:

[0078] a) Regarding airspace Based on real-time radar trajectory data and a trained LightGBM model, predict the future. Airspace traffic minutes later:

[0079] ,

[0080] in minute, For the past continuous A time slice, Represents airspace Inner past continuous The features extracted in step (2) correspond to each time slice.

[0081] b) When dynamically dividing the airspace, for each sub-airspace Based on real-time radar trajectory data and a trained LightGBM model, predict the future. Airspace traffic within minutes:

[0082] ,

[0083] in minute, For the past continuous A time slice; Represents subspace Inner past continuous The features extracted in step (2) correspond to each time slice.

[0084] This method has real-time prediction capabilities, enabling it to quickly extract features and predict traffic flow after receiving the latest radar data. This meets the timeliness requirements of dynamic airspace traffic control and significantly improves the scientific nature and response speed of air traffic control decisions.

[0085] Example:

[0086] This embodiment uses radar return track data of controlled flights from the approach / area department of a domestic airport from January to April 2025, totaling over 200,000 tracks. Each track consists of several track points, with a track point interval of 4 seconds, composed of several tags returned by radar during aircraft flight. For ease of understanding, this invention simplifies the tags used to mathematical logic descriptions, and each track includes a flight index. f timestamp t ,longitude x ,latitude y ,high z Information such as; This embodiment mainly includes the following 5 steps, and the overall flowchart can be found here. Figure 1 .

[0087] (1) Construction of spatiotemporal fusion dataset:

[0088] a) The radar trajectory reports transmitted by the aircraft during flight are referred to as the original radar trajectory dataset. Since the original dataset is trajectory data, it needs to be converted into an airspace traffic dataset. However, the sampling interval for the original dataset is 4 seconds. Directly converting it into a traffic dataset would generate a large amount of duplicate data, causing information redundancy and hindering model training. Therefore, this invention first downsamples the original radar trajectory data at the real time scale, with a downsampling interval of 1 minute. Let the original radar trajectory dataset... ,in It is a set of flight identifiers for flight trajectories. Include One sampling point, For timestamps, For three-dimensional spatial coordinates, abbreviated as .

[0089] b) Now set a fixed time slice for each downsampling time point. The trajectory data is sliced ​​into time slices with a duration of 8 seconds, resulting in a set of time slices. .

[0090] c) In each time slice Within, through statistical airspace The number of flights within a given time slot is used to represent airspace traffic. For example, to count airspace traffic at 01:01 on January 1st. For the airspace traffic within the region, select all radar trajectory reports from 01:01:00 to 01:01:08 and determine their longitude. x ,latitude y ,high z Is the altitude within the airspace? Inside, if in the airspace Internal airspace traffic +1, per flight index f It can only be recorded once within each time slice.

[0091] d) Airspace The area is divided into multiple airspace sectors, and the sector set is as follows: .

[0092] (2) Extraction of multidimensional features:

[0093] To improve the model's ability to model complex spatiotemporal relationships, multidimensional feature extraction is performed on the constructed traffic dataset, specifically for time slices. The extracted features include:

[0094] a) Time characteristics: including hourly coding Weekday Code Holiday codes The time feature uses one-hot encoding;

[0095] b) Seasonal characteristics: Month and quarter codes, with one-hot coding used for month seasonal characteristics;

[0096] c) Historical flow characteristics: past Flow sequence of time slices Recorded as ;

[0097] d) Flow statistics characteristics: The mean and variance are calculated based on a time window of 15 consecutive time slices from the past.

[0098] ,

[0099] ,

[0100] e) Flow trend characteristics: the slope of historical flow sequences,

[0101] .

[0102] (3) Construction of a traffic prediction model based on LightGBM:

[0103] a) Features extracted in step (2) As input, airspace traffic As labels, construct a supervised learning dataset.

[0104] b) The LightGBM (Light Gradient Boosting Machine) framework was used for model training and validation. Model inputs were based on past... The model uses the traffic sequence of each time slice as the primary feature, and time, season, traffic statistics, and traffic trend features as auxiliary features. The model outputs future... Predicted airspace traffic within minutes.

[0105] c) The loss function chosen is the Tweedie loss function, in this example, This loss function can effectively handle the zero-inflation and long-tail distribution characteristics of air traffic data;

[0106] d) Five-fold cross-validation (K=5) is used for model training and evaluation to ensure the robustness and generalization ability of the model.

[0107] (4) Establishment of a multi-dimensional evaluation system:

[0108] a) Construct a comprehensive model evaluation standard, using mean absolute error (MAE) as the main evaluation indicator;

[0109] b) Perform error quantile analysis to calculate the quantiles of the prediction error (e.g., 50% and 90% quantiles) to reflect the model's performance at different error levels;

[0110] c) Comparison of peak / off-peak period errors, calculating the peak period (e.g., hourly flow rate exceeding the threshold) separately. MAE during off-peak hours ensures the reliability of model predictions during critical periods;

[0111] d) Feature importance ranking: Based on the feature importance output of LightGBM, the contribution of ranked features to the prediction is determined.

[0112] e) The above multi-dimensional evaluation criteria not only comprehensively reflect the overall performance of the model, but also reveal its key performance characteristics in practical air traffic control applications, facilitating targeted optimization and decision support, and enhancing the model's interpretability and practical value. Visualization results of the evaluation indicators are available in [link to visualization]. Figures 2-6 Comparing predicted and actual values ​​( Figure 2 From the perspective of prediction error distribution, most points are distributed near the fitted line, indicating that the overall prediction effect of the model is good and can accurately reflect the actual flow trend; Figure 3 From the perspective of information gain, the error distribution is approximately symmetrical, with the center close to zero, indicating that the model as a whole does not have a systematic bias. The prediction errors are mainly concentrated in a small range, with few extreme errors, further verifying the stability and reliability of the model. Figure 4 From the perspective of [unclear], historical flow features (such as past_inside_flow, max) contribute the most to the model's prediction, while time-related features (such as hour, minute, dayofyear) and statistical features (such as mean, median) are also highly important. This indicates 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 and predicted values ​​( Figure 5 The results show that the model can capture the temporal variation of traffic flow well, and the prediction results are in high agreement with the actual observations.

[0113] During the evaluation phase, we intentionally removed most of the test samples from the early morning hours when traffic was zero, because these periods are highly predictable and the traffic is basically zero. If we retained these zero-value test samples, the overall MAE would look exceptionally good (approximately 0.55), but this would be highly misleading for practical applications. According to... Figure 6 Five-fold cross-validation curves show that the largest prediction errors occur between 10:00 and 11:00 and between 18:00 and 19:00, which are also the times of day with the highest flight traffic. This is consistent with experience and intuition, and the MAE error during these periods is about 2 to 3.

[0114] therefore,

[0115] 1. If a large number of evaluation samples with 0 values ​​are retained, the overall MAE will look particularly good (the overall MAE is about 0.55), but it cannot reflect the actual situation;

[0116] 2. If a large number of evaluation samples with zero values ​​are removed, the overall MAE will become "ugly" (the overall MAE is about 2.5), but it will better reflect the actual situation during regular flight operations. This analysis helps to discover the difference in model performance between peak and off-peak periods, providing data support for subsequent model optimization and actual air traffic control applications.

[0117] The experimental results fully verify the effectiveness and practicality of the method of this invention in real-time traffic flow prediction of air routes. The model not only has high overall prediction accuracy and a reasonable error distribution, but also can identify key features and adapt to traffic flow changes in different time periods, demonstrating promising prospects for practical application.

[0118] (5) Real-time airspace traffic prediction during the application phase:

[0119] a) In step (1), the spatial domain is divided into dynamic grids, and the grid set is set as follows: ;

[0120] b) For each airspace sector Based on real-time radar trajectory data and a trained LightGBM model (step (3)), predict the future Airspace traffic within minutes:

[0121] ,

[0122] in minute, For the past continuous A time slice;

[0123] c) This method features real-time prediction capabilities, enabling rapid feature extraction and traffic prediction upon receiving the latest radar data. This meets the timeliness requirements of dynamic airspace traffic control, significantly improving the scientific rigor and response speed of air traffic control decisions. Our invention, in conjunction with the city's air traffic control Kafka real-time trajectory and traffic system, has been deployed at this domestic airport.

Claims

1. A real-time air traffic flow prediction method based on aviation situational awareness, characterized in that... Includes the following steps: (1) Constructing a spatiotemporal fusion dataset: a) Use radar trajectory reports sent by the aircraft during flight as the original radar trajectory dataset. This includes a set of flight identifiers and the trajectory of each flight. ,in This indicates that the trajectory contains One sampling point, For timestamps, For three-dimensional spatial coordinates, abbreviated as ; b) Using fixed time slices Divide the day into time slices to obtain a set of time slices. ; c) In each time slice Within, through statistical airspace The number of flights within a given time frame represents the airspace traffic during that time segment. (2) Extraction of multidimensional features: a) Time characteristics: including hourly coding Weekday Code Holiday codes ; b) Seasonal characteristics: Month and quarter codes; c) Historical flow characteristics: past Flow sequence of time slices ; d) Traffic statistics characteristics: based on past time windows Calculate the mean and variance. ， ， e) Flow trend characteristics: the slope of historical flow sequences, ， (3) Construction of a traffic prediction model based on LightGBM: a) Using the features extracted in step (2) as input and spatial traffic as labels, construct a supervised learning dataset; The LightGBM framework was used for model training and validation. The model input was from step (2) c). The model output is the flow sequence of each time slice as the main feature, and the time, season, flow statistics and flow trend features of steps (a), (b), (d) and (e) of step (2) as auxiliary features. Predicted airspace flow within minutes , and These are hyperparameters, determined during model training; b) The Tweedie loss function is used as the loss function, defined as: ， in For the first n The actual traffic of each sample For the first n Predicted flow for each sample The total number of samples, For Tweedie distribution parameters, This loss function can effectively handle the zero-inflation and long-tail distribution characteristics of air traffic data; d) Five-fold cross-validation with K=5 is used for model training and evaluation to ensure the model's robustness and generalization ability; (4) Real-time airspace traffic prediction during the application phase: airspace Based on real-time radar trajectory data and a trained LightGBM model, predict the future. Airspace traffic minutes later: ， in minute, The number of consecutive time slices in the past. Represents airspace Inner past continuous The features extracted in step (2) correspond to each time slice.

2. The method as described in claim 1, characterized in that: The fixed time slice mentioned in step (1) It lasts for 1 minute.

3. The method as described in claim 1, characterized in that: In step (1), step c) uses the following method to calculate airspace traffic, defining airspace traffic as: ， in The indicator function is defined as follows: 。 4. The method as described in claim 1 or 3, characterized in that: Step (1) c) is followed by step d) airspace. Dynamic partitioning, from airspace Select several sub-space domains Let the grid set be ; in step (1) c) Replace with Then the subspace can be calculated. Airspace traffic within the region.

5. The method as described in claim 4, characterized in that... Step (4) is to process each subspace Based on real-time radar trajectory data and a trained LightGBM model, predict the future. Airspace traffic within minutes: ， in minute, This represents the number of consecutive time slices in the past. Represents subspace Inner past continuous The features extracted in step (2) correspond to each time slice.

Images