A data-driven hub transfer feature analysis and disengagement control method

Abstract

This invention provides a data-driven method for hub transfer feature analysis and disconnection control. It constructs an XGBoost risk prediction model with a cost-sensitive mechanism for feature analysis, and retrains a Stacking model based on the analyzed feature changes to balance precision and recall. This invention constructs a virtual connectivity network for arrivals and departures at Atlanta Airport using spatiotemporal window matching, revealing the impact mechanisms of features such as planned buffer time, arrival delays, preceding spillover effects, and hub congestion on transfer connectivity. An idealized Pareto simulation experiment is first conducted to identify the patterns for controlling transfer risks, followed by a comparative simulation experiment with a Lagrange relaxation model incorporating physical constraints. Simulation results show that this model maintains high recall while significantly reducing the overall passenger disconnection rate, providing quantitative decision support for the refined operation and management of hub airports.

Description

Technical Field

[0001] This invention belongs to the field of air transport technology, and in particular relates to a data-driven hub transfer feature analysis and disconnection control method. Background Technology

[0002] Historically, airports have been crucial nodes in integrated transportation systems. Modern large airports have evolved from simple transfer points into comprehensive transportation hubs. In air transport systems, this method of concentrating passenger flow at a hub airport and then having passengers continue their connecting flights from there is called the hub-and-spoke transportation model. This model improves the efficiency of the entire transfer system. With increasing inbound flight density, hub airports face even higher demands for passenger transfer efficiency. Furthermore, connecting flights are highly interconnected; delays in one flight often trigger ripple effects, leading to subsequent passengers experiencing transfer failures, known as passenger transfer misconnection incidents. Transfer failures at transit airports not only result in a poor customer experience but also reduce airline operational efficiency. Airlines typically address these issues by rebooking passengers. Passenger transfer misconnection incidents not only incur huge rebooking and compensation costs for airlines but also exacerbate congestion at hub airports due to passenger delays, reducing the system's overall resilience.

[0003] The industry relies on the Shortest Connecting Time (MCT) as a core indicator for ensuring successful transfers, typically set by each airline based on historical experience at different airports. However, this fixed management approach poses significant risks when facing transfer failures, as delays of the previous flight's arrival can consume a portion of the planned transfer time. This unchanging minimum connecting time fails to provide sufficient system flexibility, increasing the risk of transfer failures. However, a large number of transfers with low risk of loss of connection exist within the transfer system, and an unchanging MCT would result in significant passenger loss. Therefore, understanding how to extract transfer characteristics from large amounts of operational data and shift from fixed standards to dynamic control is crucial for improving the stable transfer capabilities of hub airports.

[0004] Existing studies have explained the operational patterns of airports and air traffic control at a macro level, but three problems remain regarding passenger operations and flight transfer control, which are the problems that this application attempts to solve using a data-driven approach.

[0005] First, hub transfer mechanisms exhibit significant nonlinear characteristics. The success of passenger transfers does not depend on a single delay threshold, but is influenced by multiple factors such as planned buffer capacity and the spillover effect of preceding flights. Traditional univariate linear analysis, relying solely on delays, struggles to effectively capture these overlooked interactive features. Second, because the proportion of lost-connection samples in the overall sample is extremely small, it easily leads to classification bias in prediction models. Lost-connection events constitute a small sample size within a large volume of operational data. If traditional regression or classification models are used, the models may overfit, prioritizing global accuracy over the majority class. However, this can result in underreporting of high-cost lost-connection events, failing to meet the high-security requirements of airports. Finally, there is a significant decision-making gap between existing transfer risk predictions and real-time airport decision-making. Existing research largely focuses on assessing delay status, lacking the logic to translate predicted probabilities into real-time resource allocation. Summary of the Invention

[0006] In view of this, this invention aims to overcome the shortcomings of the existing technology by proposing a data-driven hub transfer feature analysis and disconnection control method. Through spatiotemporal window matching technology for inbound and outbound flights, the micro-level transfer network is reconstructed, extracting multi-dimensional features such as planned buffer, arrival delay, and congestion index, and analyzing the influence mechanism of different factors on transfer connectivity. Addressing the challenge of imbalanced classification with extremely sparse disconnection samples, a class weight optimization mechanism is introduced into the XGBoost algorithm and integrated with RandomForeset to balance recall and precision. The fixed MCT management model is abandoned, and a dynamic risk scoring system is established. By comparing and analyzing the Pareto static model and the Lagrange relaxation model, the optimal and feasible transfer optimization strategy is obtained.

[0007] To achieve the above objectives, the technical solution of the present invention is implemented as follows:

[0008] A data-driven method for hub transfer characteristic analysis and disconnection control includes the following steps:

[0009] Step 1: Building a Virtual Relay Connection Network

[0010] The system acquires inbound and outbound flight operation data from hub airports, preprocesses the data, and then constructs a virtual transfer connection network between inbound and outbound flights using spatiotemporal window constraints, which include spatial constraints, carrier constraints, and time window constraints.

[0011] Step S2: Multidimensional Feature Extraction and Index Construction

[0012] Multidimensional features, including planned buffer time, arrival delay, hub congestion, and preceding ripple indicator, are extracted from the virtual transit connection network. The planned buffer time is the difference between the planned departure time of the departing flight and the planned arrival time of the arriving flight. The arrival delay is the delay duration of the arriving flight at the originating station. The hub congestion is the total number of arriving flights within the time window of the arriving flight's planned arrival. The preceding ripple indicator is a binary variable representing whether the arrival flight is delayed due to the late arrival of a preceding flight.

[0013] Step 3: Construction of a cost-sensitive discontinuity prediction model

[0014] Using the aforementioned multidimensional features as input and whether a passenger experiences a transfer loss as the output label, a classification prediction model based on the XGBoost algorithm is constructed. To address the imbalance between the proportion of lost-transfer samples and non-lost-transfer samples, a cost-sensitive learning mechanism is introduced into the XGBoost objective function. By introducing an adjustment factor into the XGBoost objective function, a greater penalty weight is applied to the classification error of positive samples.

[0015] Step 4: Integration Model Construction and Disconnection Prediction

[0016] A stacking ensemble strategy is adopted, using XGBoost and Random Forest models as base learners and logistic regression models as meta learners. The disconnection probabilities output by the base learners are used as input features of the meta learners to train the final disconnection prediction ensemble model. The optimal combination weights are learned through nonlinear mapping to predict the disconnection risk of virtual transit connections and output the final disconnection probability.

[0017] Step 5: Generation of dynamic resource allocation strategy

[0018] Based on the loss probability output by the loss prediction ensemble model and the predicted number of passengers on the virtual transit connection, a risk score is calculated for each transit connection. Based on the risk score, transit connections that need to be prioritized are selected, and a dynamic resource allocation strategy is generated.

[0019] Furthermore, in step 1,

[0020] The expression for spatial constraints is: The destination of the previous flight and the origin of the next flight are the same hub airport.

[0021] The expression for carrier constraints is: Consider connecting flights with the same airline;

[0022] The expression for the time window constraint is: , The scheduled departure time for the next flight. This refers to the scheduled arrival time of the previous flight.

[0023] Furthermore, in step 3, the cost-sensitive learning mechanism introduces an adjustment factor, scale_pos_weight, into the XGBoost objective function to impose a larger penalty weight on the classification error of positive samples. The improved weighted log loss function is defined as follows:

[0024] ;

[0025] In this formula The labels are real (1 for missing, 0 for normal). This represents the probability that the model predicts a positive sample. For positive samples, the weight coefficients are... Adaptive calculation based on the class proportions of the training set:

[0026] .

[0027] Furthermore, in step 4, the Stacking integration strategy uses 5-fold cross-validation to generate meta-features.

[0028] Furthermore, in step 4, the positive sample weight coefficient of the XGBoost model is set to 1.5 times the unbalanced ratio so that the base learner prioritizes ensuring high recall; the random forest model uses standard class-balanced weights so that the base learner focuses on ensuring high precision.

[0029] Furthermore, the formula for calculating the risk score in step 5 is as follows:

[0030] ;

[0031] in This is the disconnection probability output by the XGBoost model. This represents the number of potentially affected passengers on this transit connection.

[0032] Furthermore, the dynamic resource allocation strategy generated in step 5 specifically includes: using the Pareto detection method to select the top 15% of transit connections with the highest risk scores for priority protection, lowering the minimum transit time standard for protected transit connections, and not providing fast transit services for unprotected transit connections.

[0033] Furthermore, it also includes step 6: Lagrange relaxation optimization;

[0034] The dynamic resource allocation problem is modeled as a multi-knapsack problem with time window constraints. Non-negative Lagrange multipliers are introduced to relax the resource capacity constraint into the objective function as a penalty term. The dual problem is solved by a subgradient iterative algorithm. The resource capacity constraint is: within any time window, the number of relay connections simultaneously receiving guaranteed services shall not exceed the resource capacity limit for that time period.

[0035] Furthermore, in the Lagrange relaxation optimization, the Lagrange multiplier has the economic meaning of resource shadow price, representing the shadow price of a unit of guaranteed resource within a set time window. When the risk-reward of a flight is greater than the resource shadow price of the current period, the flight is selected for rapid transit intervention.

[0036] Furthermore, it also includes the passenger flow generation step:

[0037] The maximum number of transit passengers is calculated based on the type and number of seats of the inbound flight, the average load factor, and the transfer rate of the hub airport. The connection weight is calculated through the seat capacity of the outbound flight. A Poisson distribution is introduced to simulate the randomness of passenger arrivals. Normalization and random rounding strategies are used to ensure that the generated number of passengers meets the physical passenger capacity limit.

[0038] Compared with existing technologies, the data-driven hub transfer feature analysis and disconnection control method described in this invention has the following advantages:

[0039] This invention abandons the fixed MCT management model and establishes a dynamic risk scoring system, thereby achieving refined control over transit risks.

[0040] This invention introduces a cost-sensitive mechanism into the XGBoost objective function, setting a higher penalty weight for classification errors of missing samples by setting a positive sample weight coefficient, which significantly improves the recall rate.

[0041] This invention constructs an integrated framework based on the Stacking strategy, using cost-sensitive XGBoost (high recall) and Random Forest (high precision) as base learners, and logistic regression as a meta-learner. Meta-features are generated through 5-fold cross-validation for secondary learning, which significantly improves the accuracy of warnings while maintaining high recall, effectively solving the technical contradiction that a single model cannot balance recall and precision under imbalanced data.

[0042] This invention obtains aircraft type information by matching flight numbers with tail numbers, and uses the seat capacity of departing flights as the connection weight, giving wide-body aircraft and other large-capacity flights a higher passenger absorption weight; at the same time, it introduces Poisson distribution to simulate randomness, and uses normalization and random rounding strategies to achieve high-fidelity simulation modeling of passenger flow. Attached Figure Description

[0043] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings:

[0044] Figure 1 This is a flowchart illustrating the technical route of the method of the present invention;

[0045] Figure 2 This is a diagram showing the importance of SHAP features in this invention.

[0046] Figure 3 This is a diagram of the SHAP values ​​of the present invention;

[0047] Figure 4 This is the feature dependency graph based on SHAP values ​​of the present invention;

[0048] Figure 5 This is a diagram of the Stacking integration model architecture of the present invention;

[0049] Figure 6 This is a schematic diagram of the Pareto front curve of the present invention;

[0050] Figure 7 The resource shadow price time-varying curve of the present invention

[0051] Figure 8 This is a sensitivity analysis chart of passenger transfer rate according to the present invention. Detailed Implementation

[0052] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other.

[0053] In the description of this invention, it should be understood that the terms "center," "longitudinal," "lateral," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicating orientations or positional relationships based on the orientations or positional relationships shown in the accompanying drawings, are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, the terms "first," "second," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined with "first," "second," etc., may explicitly or implicitly include one or more of that feature. In the description of this invention, unless otherwise stated, "a plurality of" means two or more.

[0054] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art will understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0055] The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0056] refer to Figure 1 This invention provides a data-driven method for hub transfer feature analysis and disconnection control, specifically including the following:

[0057] 1. Data Feature Analysis and Indicator Construction

[0058] 1.1 Data Sources and Preprocessing

[0059] This application focuses on hub airports with high passenger transfer rates. Airports with a passenger transfer probability of over 40% are generally considered to have high transfer probability. Taking Atlanta Airport as an example, as one of the world's busiest hub airports, Atlanta's passenger transfer probability exceeds 60%. Its operational characteristics likely represent the development trend of hub airports, and its operational patterns have strong reference value for large Chinese hub airports in a period of rapid growth (such as Daxing and Pudong). To demonstrate the changes in the aviation industry structure before and after the pandemic, data from January 2019, December 2024, and January 2025 were selected, covering data from all inbound and outbound flights. Data from January 2019 and January 2025 were used in the feature analysis phase. Since the tail number and aircraft type association information for 2025 has not yet been updated, data from December 2024 was used in the simulation phase. To ensure data quality, the following cleaning logic was executed in the preprocessing phase:

[0060] (1) Outlier removal: Remove records that are canceled or missing key time fields (planned take-off and landing, actual take-off and landing).

[0061] (2) Time normalization: Convert the HHMM format timestamps of the original data into days and minutes to facilitate subsequent numerical calculations and difference analysis.

[0062] 1.2 Construction of Virtual Relay Connection Network

[0063] Raw flight data only records the operational information of a single flight. Due to passenger privacy protection, flight data lacks transit data for passenger itineraries. This application proposes a virtual connection construction method based on spatiotemporal window constraints, defining the set of inbound flights as... The set of departing flights is defined as A valid relay connection Defined as a flight pair that meets the following conditions :

[0064] The connection construction must meet the following constraints:

[0065] (1) Spatial constraints: (ATL is an abbreviation for Atlanta Hub Airport). The destination of the previous flight and the origin of the next flight are the same hub airport.

[0066] (2) Carrier constraints: This means considering connecting flights with the same airline.

[0067] (3) Time window constraint: , The scheduled departure time for the next flight. This refers to the scheduled arrival time of the previous flight. That is, the scheduled departure time of the next flight minus the scheduled arrival time of the previous flight falls within this time window.

[0068] 1.3 Definition and Physical Meaning of Key Features

[0069] To quantify the risk of connection loss during transit, this application extracts key features from three dimensions—time buffer, delay propagation, and network load—using the selection method provided by the feature selection framework. The symbolic representation and physical meaning of each indicator are shown in Table 1.

[0070] Table 1

[0071]

[0072] 1.4 Explanation of Feature Observability

[0073] To assess the feasibility of the predictive model during airport operations, a strict distinction was made between information observed at the moment of decision-making and the results of subsequent operations during the feature selection phase. The features input to the model all originated from information available before the arrival of flights, and the disconnection label was used to describe the potential risk level under a given condition. The arrival delay features input to the model were based on the departure delays of the arrival flights from their originating stations. This information is available after the flight departs. The missing flight tag, however, uses the actual arrival delay of the inbound flight. This is used to measure the actual time occupied by the delay in the planned transit buffer. This distinction in time avoids using information that is not available at the predicted time point during model training, thus preventing information leakage.

[0074] 2. Construction of a cost-sensitive discontinuity prediction model

[0075] 2.1 Criteria for Determining the Target Variable

[0076] The model first calculates the actual buffer time after the delay:

[0077] (1)

[0078] The loss of connection defined in this application refers to the potential risk of loss of connection under the assumption that subsequent flights will not experience additional delays, and is used to identify transit connections that require priority protection in advance. This refers to the actual arrival delay of inbound flights. The formula means that the planned buffer time minus the actual arrival delay of inbound flights equals the actual transfer time used by passengers.

[0079] Target variable Used to identify whether a connection loss has occurred, according to the Shortest Turnaround Time (MCT) standard, the connection loss label is defined as follows: 1 represents a connection loss, and 0 represents normal operation:

[0080] (2)

[0081] In the research scenario of this application, The minimum transfer time at the airport is 45 minutes. This transfer time is based on the minimum transfer time of Delta Air Lines, a major airline at Atlanta Airport. If a passenger's actual transfer time is less than the minimum transfer time, we consider the transfer to be a failure and an incident of passenger being left behind.

[0082] 2.2 XGBoost Algorithm Principles and Unbalanced Handling

[0083] Passenger transfer connection loss prediction is essentially a binary classification problem, which is solved by inputting the transfer connection feature vector. Predict output label (0 represents a normal connection, 1 represents a lost connection). Because the transit process involves complex relationships between planned buffers, arrival delays, and congestion, traditional linear models struggle to effectively capture the interactions between features. This application selects the XGBoost model based on gradient boosting as the classifier. This model performs well in handling large-scale tabular data and complex feature relationships. Its prediction model can be expressed as:

[0084] (3)

[0085] in, The total number of trees, For the regression tree space.

[0086] In actual aviation operations data, loss of connection is a typical low-probability event. According to statistical sample data, the ratio of positive samples (loss of connection) to negative samples (normal) is extremely disproportionate. Training directly using the standard loss function often leads to models that tend to predict all samples as normal to achieve high overall accuracy, resulting in extremely low recall for high-risk misclassification events. To address this imbalanced classification problem, this application introduces a cost-sensitive learning mechanism. By introducing a scaling factor `scale_pos_weight` into the XGBoost objective function, a larger penalty weight is applied to classification errors of positive samples. The improved weighted log loss function is defined as follows:

[0087] (4)

[0088] In this formula The labels are real (1 for missing, 0 for normal). This represents the probability that the model predicts a positive sample. These are the weighting coefficients for positive samples. Adaptive calculation based on the class proportions of the training set:

[0089] (5)

[0090] 2.3 Analysis of the Influence Mechanism of Transit Characteristics Based on SHAP

[0091] Traditional feature importance analysis can only rank the importance of individual features, but cannot reveal how features influence the model. This application introduces the SHAP interpretive analysis framework, which is based on the Shapley value concept in cooperative game theory. It can not only assess feature importance from a global perspective, but also quantify the specific value of each feature at a micro level, determining whether it promotes or inhibits the prediction of a single flight loss risk. The core idea is to construct an additive feature attribution model, interpreting the complex prediction output of the machine learning model as a linear sum of all input feature attribution values. For the prediction result of each sample, SHAP decomposes it into the sum of the baseline value and the contribution values ​​of each feature, as shown in the following mathematical expression:

[0092] (6)

[0093] in For the first The predicted risk value for each sample As a baseline model interpreter, For the first The SHAP value of each feature. This indicates that the feature increases the risk of missed connections, and vice versa. Feature importance ranking plot, SHAP value plot, and interaction plot of planned buffer time versus arrival / departure delays are shown below. Figure 2 , Figure 3 , Figure 4 As shown in the figure, the upper part of the figure represents data from January 2019, and the lower part represents data from January 2025. This application aims to find out the mechanism of feature changes before and after the epidemic by analyzing the importance of features before and after the epidemic, as well as the interaction effect between SHAP beehive graphs and features.

[0094] Figure 2 , Figure 3 This chart illustrates the changes in transit characteristics at the Atlanta hub in January 2019 and January 2025. In terms of importance, scheduled buffer time remained high in both periods, demonstrating its crucial role in ensuring successful passenger transfers. The most significant change between these two periods was the shift in arrival delays. In 2019, scheduled buffer time was far more important than arrival delays, a characteristic consistent with pre-pandemic flight operation patterns. At that time, the transit system demonstrated high resilience; given sufficient buffer time, most flights could adjust themselves.

[0095] In 2025, the importance of arrival delays surpassed planned buffers, reaching first place. This shift indicates a significant decrease in the operational efficiency of hub airports in the post-pandemic era, making transit systems more sensitive to delays in upstream flights. In terms of feature distribution, the distribution of arrival delay characteristics in 2025 was wider than in 2019, a difference due to changes in the flight operating environment. Although the total number of flights in 2025 had recovered to pre-pandemic levels, the aviation industry still faces insufficient ground support resources.

[0096] Figure 4 A feature dependency graph based on SHAP values ​​is presented. The horizontal axis represents the planned buffer time, the vertical axis represents the SHAP contribution of the feature to the risk of lost connections, and the color mapping reflects the strength of the interactive feature, arrival delay. Comparing the dependency images from January 2019 and January 2025, a significant change in the transit patterns of this hub airport can be observed, mainly in the following two aspects.

[0097] (1) The planned buffer time shows a clear negative correlation with the risk of transfer failure. The longer the buffer time, the lower the SHAP value. The inflection point where the risk approaches zero differs significantly between these two time periods. In 2019, the SHAP value converged to zero when the buffer time was around 60 minutes. This indicates that before the pandemic, about 60 minutes was sufficient to mitigate the risk of transfer failure. However, in 2025, this inflection point shifted significantly to the right, requiring approximately 80 minutes to reach zero. Furthermore, the slope in 2025 was much gentler, indicating reduced system resilience and a weakened risk mitigation effect per unit buffer time. A longer buffer time is needed to achieve the same level of safe transfers. This also demonstrates that continuing to use the fixed MCT standard is no longer suitable for the modern aviation industry's need to ensure efficient passenger transfers.

[0098] (2) The data in 2025 exhibits higher vertical dispersion. Under the same planned buffer time, the red data points with high delays in 2025 are located at higher positions and are further apart from the blue data points with low delays. In 2019, as long as sufficient planned buffer time was reserved, the arrival delays of upstream flights could be largely offset. However, in 2025, due to the decreased system resilience and reduced transfer efficiency, even with sufficient buffer time, there is still a possibility that arrival delays may result in transfer failures.

[0099] The above characteristic analysis indicates that the transit operation mode of hub airports has shifted from a resource-sufficient model before the pandemic to a resource-constrained model. Delays across the entire flight system have become more severe, and the benefits of simply increasing buffer time have significantly decreased. Moreover, continuously increasing buffer time also increases time costs for passengers. Without unlimited increases in buffer time, there is an urgent need to establish rapid transit support services for flight pairs at risk of delays.

[0100] 2.3 Selection of Evaluation Indicators

[0101] In missing-contact prediction tasks, the ratio of positive to negative samples is extremely skewed (approximately 1:9), and traditional accuracy metrics may not be the most optimal evaluation method. For example, a standard model that predicts all cases as normal can achieve 90% accuracy, but its ability to identify actual risks is very low; this is known as the accuracy paradox. Therefore, this application...

[0102] An evaluation system oriented towards business value has been established.

[0103] Given the unique nature of transit loss prediction in aviation operations, the costs associated with different types of prediction errors vary significantly:

[0104] a. False Negative (FN): The model fails to predict the flight miss, causing passengers to actually miss their flights. Consequences include rebooking compensation, accommodation arrangements, and decreased passenger satisfaction, making it extremely costly.

[0105] b. False Positives (FP): Model errors and misalignments in early warning lead to unnecessary interventions by ground crew. The main consequence is a slight redundant consumption of ground resources, with relatively manageable costs. Therefore, this application does not solely pursue accuracy but uses recall as the core evaluation metric while also considering the overall capability AUC value to assess the model. The following is the formula for calculating recall.

[0106] (7)

[0107] TP is an abbreviation for True Positive, meaning that the model correctly predicted that a passenger or a flight connection would experience a "transfer failure" (predicted as high risk), and in reality, the passenger did indeed face a transfer failure.

[0108] This model selects the following key evaluation indicators:

[0109] a. Recall Rate: A core metric. It measures the model's ability to cover real-world missing events. In the process of passengers transferring at a hub airport, a false negative means that the passenger's journey will be interrupted, but the airport has not taken any action, a situation that must be avoided at all costs.

[0110] b. Precision: A secondary indicator. It measures the accuracy of the model's warnings. A lower precision means a higher false positive rate, meaning the model predicts that passenger transfers may fail, but they do not actually fail, resulting in wasted airport resources.

[0111] c. AUC (Area Under Curve): A global metric. It measures the model's ranking ability at different thresholds. The higher the AUC, the better the model can rank high-risk flights at the top of the queue.

[0112] 2.4 Construction of the Integration Model

[0113] 2.4.1 Algorithm Framework

[0114] While the improved XGBoost model described above already possesses good interpretability, the introduction of a high penalty coefficient leads to a higher false positive rate, increasing the workload for staff. Therefore, an ensemble framework based on a stacking strategy is constructed here. This model consists of two layers: a base learning layer and a meta-learning layer. The base learning layer is constructed using XGBoost and RandomForest. XGBoost represents the Boosting paradigm, focusing on reducing model bias and uncovering deep risk features; RandomForest represents the Bagging paradigm, focusing on reducing model variance and smoothing decision boundaries. The meta-learner layer uses Logistic Regression as the meta-learner. It receives the predicted probabilities from the base models as input features, learns the optimal combination of weights through nonlinear mapping, and thus outputs the final discontinuity probability. The architecture diagram of this model is shown below. Figure 5 As shown.

[0115] 2.4.2 Base Learner Design

[0116] To achieve optimal optimization results, the base learning layer was selected from models that exhibited excellent recall and precision. In the improved XGBoost model, the weight coefficient of positive samples was set to w, which is 1.5 times the base imbalance ratio (i.e., ...). This strategy forces the model to adopt a more aggressive prediction strategy on this branch, prioritizing high recall. However, the standard Random Forest model can suppress the overfitting noise (i.e., false alarms) that may be generated by the XGBoost branch by constructing multiple decision trees through bootstrap sampling and random feature subspaces, and utilizes the Bagging ensemble property of random forests to provide robust and conservative estimates. It employs standard class-balanced weights, focusing on filtering false alarms to ensure high precision.

[0117] 2.4.3 Stacking and Fusion Strategy

[0118] Stacking fusion is not a simple voting process, but rather a secondary learning of prediction confidence. Assume the input feature vector is... The disconnection probabilities of XGBoost and Random Forest outputs are respectively and The meta-learner (Logistic Regression) constructs the following fusion function: in, For the learned fusion weights, For bias terms, The activation function is Sigmoid. Meta-features are generated using 5-fold cross-validation.

[0119] 2.4 Benchmark Comparison

[0120] This experiment used data from December 2024 to train seven models for comparative testing: standard XGBoost, improved XGBoost, standard Random Forest, improved Random Forest, improved CatBoost, improved LightGBM, and an ensemble model (improved by adopting the same weight penalty optimization mechanism) with input feature vectors. The planned buffer time, arrival delay, preceding impact indicators, and hub congestion were selected as inputs, and then the dataset was partitioned. A hold-out method was used, randomly dividing the constructed virtual connection dataset into training and test sets in an 8:2 ratio. A fixed random seed of SEED=42 was set to ensure the reproducibility of the experimental results. The main evaluation metrics are shown in Table 2.

[0121] Table 2

[0122]

[0123] 3. Simulation and Evaluation of Dynamic Resource Allocation Strategies

[0124] 3.1 Random Passenger Generation Logic Based on Capacity Weights

[0125] 3.1.1 Core Idea

[0126] Because passenger transfer information is considered personal privacy, passenger numbers are typically generated using Gaussian or Poisson distributions. Existing technologies have demonstrated the effectiveness of predictions based on these distributions in practice. However, relying solely on virtual distributions is insufficient. This application establishes a correspondence between flights and their corresponding aircraft types by mapping flights to their tail numbers. The maximum number of transfer passengers is calculated using inbound flights, and the transfer weight is calculated using the ratio of the number of seats on outbound flights to the sum of all connecting seats. The simulation method here does not evenly distribute passenger traffic across every virtual connection but rather simulates the traffic differences between different transfer connections at a hub airport. Specifically, it is assumed that the aircraft type capacity of outbound flights represents different transfer ratios, meaning that larger aircraft (such as wide-body aircraft) have a higher passenger absorption weight than smaller aircraft (such as regional jets). Simultaneously, a Poisson distribution is introduced to simulate the randomness of passenger arrivals, and capacity constraint normalization is applied to ensure that the generated number of passengers strictly conforms to physical passenger capacity limits.

[0127] 3.1.2 Mathematical Model Description

[0128] The process of constructing passenger flow data is as follows:

[0129] (a) First, determine the maximum total number of transit passengers for each inbound flight. This application assumes that the maximum total number of transit passengers is determined by the number of seats on the aircraft type, the average load factor, and the transit rate of the hub airport. The mathematical expression can be represented as:

[0130] (8)

[0131] For the maximum number of transit passengers, This refers to the number of seats corresponding to this aircraft model. The average occupancy rate, This represents the transfer rate at this hub airport. Based on the characteristics of Atlanta hub airport, The value was selected as 0.85. The value was selected as 0.6.

[0132] (b) For inbound flights Each departing flight connected within the valid transit window We define its connection weights Equivalent to departing flights Seating capacity Passengers then flow to flights. Allocation probability for:

[0133] (9)

[0134] in For flights The set of all feasible outbound connections, here This represents the total number of seats on all departing flights. This weighting mechanism ensures that larger aircraft have a greater capacity to accommodate passengers and also ensures that traffic flow is tilted towards wide-body aircraft and popular routes.

[0135] (c) To simulate the random fluctuations in passenger numbers in the real world, a Poisson distribution is introduced here, with specific connections... Passenger numbers Follows a Poisson distribution:

[0136] (10)

[0137] The number of passengers on this transit route will fluctuate as the maximum number of passengers is multiplied by the weight of the route.

[0138] (d) After introducing the Poisson distribution, the number of passengers generated here may occasionally exceed the maximum total number of passengers for inbound flights. To address this overflow phenomenon and ensure the realism of the simulation, this application introduces a normalization solution, with the normalization coefficient being:

[0139] (11)

[0140] Here from arrive The total number of people generated by all transit connections is The physical passenger capacity of inbound flights is The normalization factor here ensures that the number of generated passengers will only be scaled down proportionally if the number exceeds the limit; otherwise, it will remain unchanged. When the number of generated passengers exceeds the limit, a scaling factor is applied to obtain the corrected number of passengers:

[0141] (12)

[0142] After multiplying by the correction factor, the result will most likely have a decimal part, so we use a random rounding strategy here. Decompose into integer parts and decimal part The final integer number of passengers. The following is confirmed:

[0143] (13) This operation involves introducing a uniformly distributed random variable. Implementation, that is, when Round up to the nearest integer.

[0144] This rounding strategy can further improve simulation accuracy. For example, if the calculated number is 0.2 people, there is a 20% probability that it will become 1 person. This strategy is very important in simulations, as even a delay of 1 person can be captured by the model, greatly enhancing the model's sensitivity to small numbers of people transferring between tasks.

[0145] 3.2 Model Evaluation Indicators

[0146] To comprehensively evaluate the effectiveness of different strategies, this application introduces the following two core metrics:

[0147] (1) Total number of passengers rescued

[0148] This metric reflects the number of passengers whose travel plans were not disrupted after the strategy was implemented. The number of passengers who lost contact under the baseline scenario. The total number of passengers lost to the vehicle after the strategy was improved.

[0149] (14)

[0150] (2) Improvement rate of disconnection

[0151] This metric reflects the stability of airport operational capacity and the resilience of the overall network. It eliminates the influence of traffic volume, facilitating cross-model comparisons.

[0152] (15)

[0153] In the actual operation of hub airports, the proportion of flights that can provide fast transfers is limited, so we cannot provide fast transfer services for all connecting flights.

[0154] Simply relying on predicted probabilities for control may have limitations. Therefore, this application also introduces the number of transit passengers. Flights with a high probability of being out of service may carry only a very small number of transit passengers, while wide-body aircraft flights with a moderate probability of being out of service may carry a large number of transit passengers. To address this trade-off, this application defines transit connections. Risk score Expected value for potential passenger loss:

[0155] (16)

[0156] The disconnection probability is output by the XGBoost model, reflecting the likelihood of the risk occurring. This represents the number of potentially affected passengers on the connection, reflecting the severity of the risk.

[0157] 3.3 Pareto Detection and Analysis

[0158] The Pareto test used here does not consider real-world conflict constraints; its aim is to first find a theoretically optimal upper bound, which will then be compared with the Lagrange relaxation optimization model below after introducing conflict constraints.

[0159] The maximum amount of resources available for airport access is set at 50%. The objective of this test is to identify an optimal set of interventions. This maximizes the expected sum of risks associated with virtual connections within the set. This strategy employs an efficient sorting and truncation method:

[0160] (17)

[0161] in This represents the total number of virtual relay connections. (Parameter) The selection of airports determines the proportion of fast transfer services that an airport can provide. To establish... To determine the optimal value, this experiment uses the December 2024 test set data to design an ergodic simulation experiment, and constructs an analytical model through Pareto experiments. The resource input ratio is set. The calculation is performed within the interval with a step size of 0.01. For each... The value is used to calculate the corresponding total revenue (i.e., the total number of passengers saved). The simulation results generated are as follows: Figure 6The Pareto front curve is shown. This curve exhibits a significant characteristic of a continuously decreasing slope: when... At that point, the curve slope was extremely steep, indicating that prioritizing and optimizing a few high-risk head flights could lead to a significant improvement in transfer efficiency. When the curve flattens out, it indicates that increasing resource input at this point will not significantly improve the number of passengers saved by the system.

[0162] To find an efficiency balance between maximizing the number of passengers recovered and minimizing resource consumption, an inflection point detection algorithm based on the maximum distance method is introduced. The detection function is defined as follows: :

[0163] (18)

[0164] Through calculation The maximum point was found, and the theoretical optimal solution was determined in the simulation experiment. 19.0. At this point, the unit resource guarantee capability reaches its optimal performance. Although the theoretical inflection point is 19%, a certain engineering margin is allocated here for system stability to cope with unexpected situations in the model. Therefore, =15% is the optimal interference ratio derived from data-driven simulation under ideal conditions.

[0165] To verify the above strategy, this application simulates the priority-based protection of the physical protection process in a simulation environment. Based on logical judgment... The system executes the following dual-track MCT standard: (19)

[0166] If the risk score of the transit route is in the top 15%, express transit service will be provided, reducing the minimum transit time to 25 minutes; if the risk score is not in the top 15%, express transit service will not be provided.

[0167] 3.4 Lagrange Relaxation Optimization Model

[0168] While the Pareto experiment calculates the optimal coverage range, it is essentially a physical cutoff and does not take into account conflicts in real-world physical resources. During peak airport operations, high-risk flights may suddenly increase in number. If coverage is based solely on the calculated risk score, resource demands (such as shuttle buses and express services) during a specific period may far exceed the service capacity available at that time, causing the strategy to fail.

[0169] To verify whether a good optimization effect can still be achieved after introducing real-world physical constraints, the dynamic resource allocation problem here is modeled as a multi-knapsack problem with time window constraints.

[0170] 3.4.1 Problem Definition and Mathematical Model

[0171] set up Given a set of discrete time slices representing the entire day's running time, a day is divided into... A length of (This application takes a 15-minute time window, i.e.) .set up This is a pool of candidate high-risk flights arriving in Hong Kong that day. For any flight... Define the time window in which its arrival time falls as .

[0172] Introducing 0-1 decision variables :

[0173]

[0174] (1) Model objective function

[0175] The optimization objective of this model is to maximize the total expected revenue saved by the entire system for transit passengers.

[0176] (20)

[0177] in, For flights The risk-reward ratio is defined as the product of the probability of the flight losing connection and the number of passengers potentially losing connection. .

[0178] (2) Physical constraints

[0179] The model must satisfy physical resource constraints, i.e., within any time window. Within a given timeframe, the number of relay connections simultaneously receiving support services must not exceed the resource capacity limit for that period. (This application takes 40):

[0180] (twenty one)

[0181] (twenty two)

[0182] Among them, constraint (21) restricts resource consumption within a single time window. Since flight arrival times are not uniform, simple global sorting strategies (such as the Pareto principle) are prone to failure under this constraint, so an optimization algorithm is needed for dynamic allocation. This will cause the decisions of different flights to be interrelated, making the problem difficult to solve directly.

[0183] 3.4.2 Lagrange Relaxation and Duality Problem

[0184] To decouple the above constraints, a non-negative Lagrange multiplier vector is introduced. The constraint (21) is relaxed into the objective function as a penalty term to construct the Lagrangian function. :

[0185] (twenty three)

[0186] Reorganizing formula (23) by combining the summation terms by individual flight, we get:

[0187] (twenty four)

[0188] Thus, the original global optimization problem is transformed into finding the dual problem. The optimal solution:

[0189] (25)

[0190] 3.4.3 Economic Explanation

[0191] In formula (24), the resource shadow price is... It has a clear economic meaning. Represents a time window The shadow price of resources guaranteed by the unit, i.e., opportunity cost. When resources are extremely scarce at a certain time, the corresponding multiplier... It will rise, only when the flight Risk and return Greater than the resource cost of the current period Time (i.e., net utility) Only then will the system recommend intervention. This, through the price mechanism, curbs resource requests from low-value flights during congested periods.

[0192] 3.4.4 Subgradient-based solution algorithm

[0193] because It is about The function is convex but not differentiable. This application uses the subgradient method for iterative solution. The algorithm flow is as follows:

[0194] Step 1: Set the initial multipliers Set the maximum number of iterations. and initial step size .

[0195] Step 2: In the In the next iteration, based on the current multiplier For each flight Independent decision-making:

[0196] (26)

[0197] Step 3: Since the relaxed solution may violate the original capacity constraint (21), a feasibility repair is required. For any time window that violates the constraint... (Right now ),according to Sort by size from largest to smallest, keeping the first few largest. One flight The rest are forcibly set to 0, thus obtaining a feasible solution. .

[0198] Step 4: Calculate the subgradient :

[0199] (27)

[0200] Update the multiplier vector:

[0201] (28)

[0202] Where step size Decrease with iteration count (e.g.) This is to ensure that the algorithm converges.

[0203] Step 5: If the maximum number of iterations is reached or the duality gap is less than the threshold, output the historically best feasible solution. .

[0204] 3.5 Comparative Analysis of Experiments on the Two Models

[0205] In the Lagrange relaxation model, the upper limit of physical resource capacity This is a key parameter determining optimization capability. We do not use specific physical resources such as shuttle buses here to ensure better adaptability of the model. The capacity of each airport may differ; in this experiment, it is set based on the actual operational scale of Atlanta Airport, approximately 70% of peak demand, chosen as 40. During off-peak periods, resources are abundant (…). During peak periods, demand exceeds capacity, forcing the model to exclude some requests based on shadow prices. To verify the applicability of the Lagrange and Pareto models in real-world environments, this application conducts comparative experiments on the same dataset using a dynamic resource allocation strategy based on Lagrange relaxation and a traditional static Pareto truncation strategy. The experiments are based on a test set from December 2024, which constructed 204,847 connections. Without any strategy, the number of disconnected connections was 44,532. Table 3 shows the simulation results for the two strategies.

[0206] Table 3

[0207]

[0208] Table 3 shows that the two models achieve similar levels of improvement, but their underlying resource allocation methods are completely different. The Pareto model statically selects the top 15% of flights. During peak arrival periods, these 15% of flights are highly likely to arrive simultaneously, causing demand to exceed resource limits and preventing rapid support services from being implemented. In contrast, the Lagrange relaxation model is based on the maximum support services that can be provided within a time slice, making it more feasible. Furthermore, although the improvement rates are similar, the intervention ratios differ, demonstrating that the Pareto model is based on ideal conditions. To achieve the ideal Pareto effect, the intervention ratio needs to be significantly increased. This also clearly demonstrates the feasibility of the Lagrange relaxation model.

[0209] like Figure 7 As shown, this is a time-varying curve of resource shadow prices as of December 1, 2024.

[0210] The vertical axis in the figure The shadow price of resources guaranteed by representative units reflects the marginal opportunity cost of resources within a specific time window. When The time frame (e.g., 10.0-11.5 hours) indicates that airport ground support resources are abundant, the marginal cost of all intervention requests is zero, and the system exhibits a state of resource surplus. The graph shows several distinct peaks, corresponding to the peak periods of high concentration of inbound flights at the hub airport. During these periods, resource demand exceeds... The physical upper limit. The surge in shadow prices has an automatic adjustment function, only when the risk-reward of the flight... Shadow price beyond that moment Only when a flight is selected by the system for rapid transfer intervention will it be eligible. The peak value is close to 0.7, meaning that during extremely congested periods, each additional intervention slot will place a very high burden on the system. Through this dynamic pricing mechanism, the model effectively suppresses the occupation of resources by low-value flights during congested periods, and instead allocates limited resources such as shuttle buses and near-gate positions to flights with the greatest potential for rescue, thereby maximizing the overall rescue efficiency of the system without increasing the total amount of resources.

[0211] 3.6 Sensitivity Analysis of Passenger Transfer Rate at Hub Airports

[0212] To verify whether the transfer rate of arriving flight passengers would have a significant impact on the control system after the introduction of the Lagrange relaxation model, a sensitivity experiment was conducted. The fixed 60% transfer rate of hub airport passengers was changed to 30%, 40%, 50%, 60%, 70%, and 80%, and the experiment was repeated to obtain the number of passengers saved and the proportion of transfer passengers under different passenger transfer scenarios at the hub airport.

[0213] like Figure 8As shown, after introducing this method, the number of passengers saved increases with the proportion of passengers transferring flights, and the improvement rate remains high and stable under various transfer rates, indicating that the system has high stability.

[0214] 3.7 Economic Value Analysis

[0215] In actual airport service scenarios, we should not only consider the statistical indicators of predictive models, but also the costs in the real-world environment. The scenario studied in this application exhibits typical imbalance costs, where the cost of passenger mishandling far exceeds the cost of a rapid service. Therefore, we introduce a cost-benefit analysis model. Let... The comprehensive losses caused by a single passenger mis-boarding incident include passenger rebooking fees, compensation for food and accommodation, subsequent seat occupation, and potential damage to brand reputation. The redundant costs incurred by a single false alarm mainly include the ineffective attendance of ground service personnel or the fuel costs of shuttle vehicles. In actual operation, this typically meets... Based on industry experience and data estimates, the comprehensive compensation cost for a single international transit passenger experiencing a loss of connection is... The cost is approximately 3,000 yuan (including rebooking and compensation), while the cost of a single ground rapid guidance is... It costs only about 50 yuan (including labor and energy consumption). Based on the performance data (accuracy) of the improved ensemble model in this application... This means that every time the model issues a signal... Of the 16 warnings, only one was a genuine risk, while the remaining 0.16 were redundant safeguards.

[0216] At this point, the system's expected net return It can be represented as:

[0217] (29)

[0218] Substitute the above estimated values:

[0219] (30)

[0220] The above calculations show that although the lower accuracy of the improved model results in approximately 0.16 times more redundancy, this strategy of sacrificing redundancy for higher returns is economically highly beneficial compared to the extremely costly cost of rescuing out-of-line events. Therefore, the decision-making logic proposed in this study, which sacrifices some accuracy for high recall, is robust and economically reasonable in engineering practice.

[0221] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A data-driven method for hub transfer characteristic analysis and disconnection control, characterized in that: Includes the following steps: Step 1: Building a Virtual Relay Connection Network The system acquires inbound and outbound flight operation data from hub airports, preprocesses the data, and then constructs a virtual transfer connection network between inbound and outbound flights using spatiotemporal window constraints, which include spatial constraints, carrier constraints, and time window constraints. Step S2: Multidimensional Feature Extraction and Index Construction Multidimensional features, including planned buffer time, arrival delay, hub congestion, and preceding ripple indicator, are extracted from the virtual transit connection network. The planned buffer time is the difference between the planned departure time of the departing flight and the planned arrival time of the arriving flight. The arrival delay is the delay duration of the arriving flight at the originating station. The hub congestion is the total number of arriving flights within the time window of the arriving flight's planned arrival. The preceding ripple indicator is a binary variable representing whether the arrival flight is delayed due to the late arrival of a preceding flight. Step 3: Construction of a cost-sensitive discontinuity prediction model Using the aforementioned multidimensional features as input and whether a passenger experiences a transfer loss as the output label, a classification prediction model based on the XGBoost algorithm is constructed. To address the imbalance between the proportion of lost-transfer samples and non-lost-transfer samples, a cost-sensitive learning mechanism is introduced into the XGBoost objective function. By introducing an adjustment factor into the XGBoost objective function, a greater penalty weight is applied to the classification error of positive samples. Step 4: Integration Model Construction and Disconnection Prediction A stacking ensemble strategy is adopted, using XGBoost and Random Forest models as base learners and logistic regression models as meta learners. The disconnection probabilities output by the base learners are used as input features of the meta learners to train the final disconnection prediction ensemble model. The optimal combination weights are learned through nonlinear mapping to predict the disconnection risk of virtual transit connections and output the final disconnection probability. Step 5: Generation of dynamic resource allocation strategy Based on the loss probability output by the loss prediction ensemble model and the predicted number of passengers on the virtual transit connection, a risk score is calculated for each transit connection. Based on the risk score, transit connections that need to be prioritized are selected, and a dynamic resource allocation strategy is generated.

2. The data-driven hub transfer characteristic analysis and disconnection control method according to claim 1, characterized in that: In step 1, The expression for spatial constraints is: The destination of the previous flight and the origin of the next flight are the same hub airport. The expression for carrier constraints is: Consider connecting flights with the same airline; The expression for the time window constraint is: , The scheduled departure time for the next flight. This refers to the scheduled arrival time of the previous flight.

3. The data-driven hub transfer characteristic analysis and disconnection control method according to claim 1, characterized in that: In step 3, the cost-sensitive learning mechanism introduces a scaling factor, scale_pos_weight, into the XGBoost objective function to impose a larger penalty weight on the classification error of positive samples. The improved weighted log loss function is defined as follows: ； In this formula The labels are real (1 for missing, 0 for normal). This represents the probability that the model predicts a positive sample. For positive samples, the weight coefficients are... Adaptive calculation based on the class proportions of the training set: 。 4. The data-driven hub transfer characteristic analysis and disconnection control method according to claim 1, characterized in that: In step 4, the Stacking integration strategy uses 5-fold cross-validation to generate meta-features.

5. The data-driven hub transfer characteristic analysis and disconnection control method according to claim 1, characterized in that: In step 4, the positive sample weight coefficient of the XGBoost model is set to 1.5 times the unbalanced ratio so that the base learner prioritizes ensuring high recall; the random forest model uses standard class-balanced weights so that the base learner focuses on ensuring high precision.

6. The data-driven hub transfer characteristic analysis and disconnection control method according to claim 1, characterized in that: The formula for calculating the risk score in step 5 is as follows: ； in This is the disconnection probability output by the XGBoost model. This represents the number of potentially affected passengers on this transit connection.

7. The data-driven hub transfer characteristic analysis and disconnection control method according to claim 1, characterized in that: The dynamic resource allocation strategy generated in step 5 specifically includes: using the Pareto detection method to select the top 15% of transit connections with the highest risk scores for priority protection, lowering the minimum transit time standard for protected transit connections, and not providing fast transit services for unprotected transit connections.

8. The data-driven hub transfer characteristic analysis and disconnection control method according to claim 1, characterized in that: It also includes step 6: Lagrange relaxation optimization; The dynamic resource allocation problem is modeled as a multi-knapsack problem with time window constraints. Non-negative Lagrange multipliers are introduced to relax the resource capacity constraint into the objective function as a penalty term. The dual problem is solved by a subgradient iterative algorithm. The resource capacity constraint is: within any time window, the number of relay connections simultaneously receiving guaranteed services shall not exceed the resource capacity limit for that time period.

9. The data-driven hub transfer characteristic analysis and disconnection control method according to claim 8, characterized in that: In the Lagrange relaxation optimization, the Lagrange multiplier has an economic meaning of resource shadow price, representing the shadow price of a unit of guaranteed resource within a set time window. When the risk-reward of a flight is greater than the resource shadow price in the current period, the flight is selected for rapid transit intervention.

10. The data-driven hub transfer characteristic analysis and disconnection control method according to claim 1, characterized in that: It also includes the passenger flow generation step: The maximum number of transit passengers is calculated based on the type and number of seats of the inbound flight, the average load factor, and the transfer rate of the hub airport. The connection weight is calculated through the seat capacity of the outbound flight. A Poisson distribution is introduced to simulate the randomness of passenger arrivals. Normalization and random rounding strategies are used to ensure that the generated number of passengers meets the physical passenger capacity limit.

Images