Trajectory-driven aircraft taxiing planning method and system for airport surfaces

Abstract

This invention discloses a trajectory-driven aircraft taxiing planning method and system for airport surfaces. The method includes: performing topology modeling based on the physical structure of the airport surface and introducing the node2vec algorithm to generate surface structure feature vectors; constructing a taxiing trajectory temporal prediction model combining a bidirectional long short-term memory network and an attention mechanism for individual aircraft taxiing trajectories to achieve accurate prediction of individual aircraft taxiing trajectories; establishing a spatiotemporal collaborative two-layer planning model for aircraft group taxiing performance, where the upper-layer model optimizes runway utilization efficiency and the lower-layer model optimizes taxiing operation efficiency; and optimizing and solving the two-layer planning model to generate an optimized scheme that balances taxiing efficiency and runway operation efficiency. This invention effectively solves the problem of insufficient accuracy in traditional taxiing trajectory prediction and can be used for ordered processes such as surface performance optimization, effectively reducing total taxiing time, runway occupancy time, and delay time.

Description

Technical Field

[0001] This invention relates to a trajectory-driven airport surface aircraft taxiing planning method and system, particularly an intelligent planning method and system based on airport surface structural characteristics, aircraft taxiing trajectory prediction, and group taxiing performance optimization, belonging to the field of intelligent airport surface management. Background Technology

[0002] The long taxiing times, limited conflict resolution solutions, and heavy workloads and error-prone nature of air traffic controllers have become key bottlenecks restricting airport traffic flow and operational efficiency. To address the challenges of Air Traffic Management (ATM) systems, developed aviation nations have proposed the Trajectory-Based Operation (TBO) concept, which uses time control to precisely monitor aircraft tracks, thereby improving operational efficiency and safety. However, with the deepening application of TBO, this concept is gradually extending to the intelligent management of surface operations, especially in complex airport scenarios. Precise taxiing trajectory prediction and planning facilitate efficient control of surface aircraft taxiing paths. Against this backdrop, the need for intelligent operation management of surface aircraft is becoming increasingly urgent. To effectively address these challenges, airports, air traffic control, and airlines are actively exploring intelligent and automated solutions to achieve precise allocation of critical resources. Currently, there is significant room for improvement in the intelligent management of aircraft at busy airports in my country, mainly in two aspects: first, to accurately predict aircraft taxiing trajectories under complex airport structures in order to optimize trajectories; and second, to optimize the overall performance of "group" operations based on individual trajectory predictions and combined with airport taxiing performance indicators, thereby improving airport operational efficiency and safety.

[0003] Therefore, the big data-driven taxiway trajectory prediction and optimization control technology proposed in this invention can not only improve the intelligence and precision of surface control, refine the management and allocation of overall resources, and optimize aircraft operating patterns, but also provide controllers with a unified view of aircraft positions and pre-planned resource scheduling, thereby improving airport collaborative decision-making. Simultaneously, accurately predicting the taxiways of surface aircraft allows controllers to identify and optimize potential trajectory conflicts in advance, significantly reducing the cognitive load on control personnel and helping them allocate time and resources more effectively according to the risk levels of different areas, thus significantly improving airport surface control efficiency. In this context, intelligently optimizing taxiways and key node times through scientific control methods can not only reduce fuel consumption during taxiing but also significantly improve taxiing safety, becoming a core solution for improving airport resource utilization and alleviating surface congestion. Summary of the Invention

[0004] The purpose of this invention is to provide a trajectory-driven airport surface aircraft taxiing planning method and system, aiming to solve problems such as busy surface traffic, decreased operational efficiency, and rising flight delay rates caused by the increase in airport traffic. By constructing an intelligent planning method and system based on airport surface structural characteristics and taxiing trajectory prediction, it can achieve accurate prediction of individual aircraft taxiing trajectories and optimized scheduling of group taxiing performance, thereby improving surface operational efficiency, reducing taxiing conflicts and delays, and providing scientific support for the intelligent and precise management and control of airport surface operations.

[0005] To achieve the above objectives, the present invention adopts the following technical solution:

[0006] A trajectory-driven airport surface aircraft taxiing planning method includes the following steps:

[0007] Step 1: Perform topological modeling based on the physical structure of the airport surface and introduce the node2vec algorithm to generate surface structure feature vectors; combine historical taxiing trajectory data to extract multi-dimensional feature vectors of aircraft and construct a comprehensive feature representation of the aircraft.

[0008] Step 2: For the individual taxiing trajectory of an aircraft, based on the multivariate features generated in Step 1, construct a taxiing trajectory temporal prediction model (Bi-LSTM-A) that combines a bidirectional long short-term memory network with an attention mechanism to achieve accurate prediction of the individual taxiing trajectory of the aircraft.

[0009] Step 3: For the taxiing performance of aircraft "group", based on the accurate trajectory prediction dataset, establish a two-layer planning model for surface taxiing spatiotemporal coordination, in which the upper-layer model optimizes runway utilization efficiency and the lower-layer model optimizes taxiing operation efficiency.

[0010] Step 4: The non-dominated sorting multi-objective genetic algorithm (NSGA-II) is used to optimize and solve the bi-level programming model, generating an optimized scheme that takes into account both taxiing efficiency and runway operation efficiency.

[0011] Furthermore, step 1 includes the following sub-steps:

[0012] Step 11: Perform topological modeling of the physical structure of the airport surface. By simplifying the complex structure of the airport surface and highlighting the layout characteristics of different functional areas within the surface, the complexity of the model can be reduced.

[0013] Step 12: The node2vec node embedding algorithm is used to capture the overall structure of the scene structure graph and the local environment of the sliding nodes, preserve the similarity between nodes, and generate scene structure feature vectors.

[0014] Step 13: Use one-hot coding to finely mark special nodes, including parking positions, runway entrances and exits, thereby improving the accuracy of taxiing behavior prediction for such structures;

[0015] Step 14: Combine historical taxiing trajectory data to extract multi-dimensional feature vectors of the aircraft, including taxiing speed, acceleration, and latitude and longitude coordinates, and construct a comprehensive feature representation of the aircraft.

[0016] Step 15: Combine the scene structure feature vector generated in Step 13 with the aircraft multidimensional feature vector constructed in Step 14 to form a complete feature input, providing a data foundation for subsequent taxiing trajectory prediction.

[0017] Furthermore, step 12 is detailed as follows:

[0018] Graph adjacency matrix constructed from the scene mixing graph The entire scene structure is abstracted into a maximum likelihood optimization problem using node2vec; where As a mapping function from taxiway nodes to their corresponding feature representations, For size The matrix is ​​s, where s is the dimension of the feature representation, and for each taxiway node... Then its neighborhood node set Similarly satisfied ,in For the local node set of the taxiway, the subscript is... For taxiway nodes The chosen sampling method aims to maximize the neighborhood log probability density function of all taxiway nodes, as shown in the following expression:

[0019]

[0020] To facilitate optimization, the following assumptions are introduced:

[0021] (1) Neighborhood Independence: For each taxiway node, the probability of exploring its corresponding neighboring node is independent of other neighboring nodes, as shown in the following expression:

[0022]

[0023] (2) Feature symmetry: For each taxiway node, its feature representation as a neighbor of other taxiway nodes is the same as its feature representation in its own space, as shown in the following expression:

[0024]

[0025] in, This indicates the taxiway node. taxiway node The probability of selection The softmax function is used to represent the function for taxiway nodes. The overall probability of selecting all other taxiway nodes;

[0026] Based on the above assumptions, the complex formula is simplified as follows:

[0027]

[0028] in, Because of the excessively high computational cost, the multi-classification problem is transformed into multiple binary classification problems for resampling by introducing negative sampling.

[0029] A random walk strategy is employed to capture the structural relationships and spatial connections between scene nodes, starting from the glide path nodes within a random walk. Jump to the taxiway node The conditional probability is expressed as follows:

[0030]

[0031] in For the th random walk sequence 1 node Indicates from the taxiway node Jump to the taxiway node The nonnormalized probability is determined by the edge weight. It is a normalization constant obtained through negative sampling;

[0032] A random walk strategy based on a second-order Markov chain is employed, where the selection of the next taxiway node is influenced by both the current and previous taxiway nodes. The bias of the random walk is controlled by adjusting two parameters, emphasizing depth-first search (DFS) and breadth-first search (BFS) respectively, thus employing taxiway node walk preference parameters. According to the search strategy, the roaming path just past the glide path node... Jump to the taxiway node From the current point Start moving to the next point The walk, expressed as follows:

[0033]

[0034] in Represents taxiway node and The shortest distance between elements is independent of the distance weight of the edges; The return parameter represents the probability that the traverse will return to a previously visited gliding path node. If so, a breadth-first search strategy is adopted; if If so, a depth-first search strategy is adopted; The in / out parameters represent the breadth of the wander direction. Then a breadth-first search strategy is adopted, which walks through already explored structures. Then, a depth search strategy is adopted to walk into unexplored structures.

[0035] Furthermore, step 2 includes the following sub-steps:

[0036] Step 21: Conduct in-depth analysis of the individual taxiing trajectories of aircraft in the airport area, and establish a time series prediction model based on the constructed multi-dimensional characteristics of aircraft taxiing trajectories;

[0037] Step 22: Construct a gliding trajectory temporal prediction model (Bi-LSTM-A) based on a bidirectional long short-term memory network (Bi-LSTM) and an attention mechanism to simultaneously capture the temporal relationship and spatial features between gliding trajectory points;

[0038] Step 23: The ROUGE-N evaluation system, root mean square error (RMSE), and mean absolute error (MAE) are used as performance evaluation indicators to evaluate the prediction accuracy of the taxiing trajectory time-series prediction (Bi-LSTM-A) model.

[0039] Step 24: Through verification analysis of typical taxiing trajectories and taxiing times, evaluate the prediction accuracy of the model and output the accurate predicted taxiing trajectory of the aircraft "individual".

[0040] Furthermore, the specific indicator design for step 23 is as follows:

[0041] (1) Root Mean Square Error (RMSE)

[0042]

[0043] (2) Mean Absolute Error (MAE)

[0044]

[0045] In the above two formulas, For the first The true value of each sample For the first Predicted values ​​for each sample;

[0046] (3) ROUGE-N evaluation index system (N-gram Co-Occurrence Statistics, ROUGE-N)

[0047]

[0048]

[0049]

[0050] in This represents the number of n-points in the predicted trajectory that match the sample trajectory. This represents the number of n-points present in the sample trajectory. This represents the number of n-points present in the predicted trajectory.

[0051] Furthermore, step 3 includes the following sub-steps:

[0052] Step 31: Based on the accurate taxiing trajectory prediction dataset and taxiing performance index system, evaluate taxiing performance under different traffic scenarios to provide data support for optimizing scheduling;

[0053] Step 32: Establish a two-layer planning model for spatiotemporal coordination of surface gliding based on accurate trajectory prediction, including an upper-layer model and a lower-layer model;

[0054] Step 33: The relevant sets, parameters, and variables required by the upper-level model are shown below:

[0055] variable:

[0056] If the flight Prior to flights Start occupying the runway but ,otherwise ;

[0057] If the flight The runway selection is but ,otherwise ;

[0058] :flight On the runway Planned take-off and landing times;

[0059] parameter:

[0060] Number is Flights;

[0061] Approaching flights On the flight The wake spacing ahead;

[0062] :track and The relevant approach slant ranges between them;

[0063] :flight The final approach velocity;

[0064] :flight The time the runway occupancy time;

[0065] :flight On the runway The planned takeoff and landing times;

[0066] The earliest time in the current planning period;

[0067] :track Start time of use;

[0068] :track End time of use;

[0069] :track Conflict time of use;

[0070] :flight The maximum acceptable delay time;

[0071] gather:

[0072] : The scheduled time for the assembly of flights, among which ,at the same time This indicates that there are a total of [number] people on the field within the planned time. Flights;

[0073] Arrival of flights within the planned timeframe;

[0074] Assembly point for departing flights within the planned timeframe;

[0075] The set of runways to choose from, among which ;

[0076] Airport surface taxiway node collection, among which ;

[0077] The set of nodes of the airport surface runway;

[0078] Airport surface parking stand node collection;

[0079] :flight Heading to the runway The set of nodes for the predicted gliding path;

[0080] :flight Heading to the runway The set of nodes for the shortest glide path;

[0081] :flight Heading to the runway The set of time points that have passed;

[0082] :flight Heading to the runway The set of time points that have passed;

[0083] Step 34: Construct the objective function of the upper-level model, and consider runway selection constraints, runway spacing constraints, runway occupancy constraints, optimization range constraints, and maximum delay constraints.

[0084] Step 35, the relevant sets, parameters and variables required by the lower-level model are as follows:

[0085] variable:

[0086] :flight go through The node's time, if not elapsed ;

[0087] If the flight The path passes through the taxiway node. but ,otherwise ;

[0088] If the flight Prior to flights Crossing the taxiway node but ,otherwise ;

[0089] parameter:

[0090] Number is taxiway nodes;

[0091] :flight and Minimum safe taxiing distance between them;

[0092] taxiway node and The minimum distance between them;

[0093] :flight Passing through the taxiway node speed;

[0094] Minimum turnaround time for flights;

[0095] : Apron occupancy time for flights;

[0096] Departure flights The planned launch time;

[0097] Number of skiing incidents during the planned time period;

[0098] : Maximum constant value;

[0099] gather:

[0100] : The scheduled time for the assembly of flights, among which ,at the same time This indicates that there are a total of [number] people on the field within the planned time. Flights;

[0101] Arrival of flights within the planned timeframe;

[0102] Assembly point for departing flights within the planned timeframe;

[0103] The set of runways to choose from, among which ;

[0104] Airport surface taxiway node collection, among which ;

[0105] The set of nodes of the airport surface runway;

[0106] Airport surface parking stand node collection;

[0107] :flight Heading to the runway The set of nodes for the predicted gliding path;

[0108] :flight Heading to the runway The set of nodes for the shortest glide path;

[0109] :flight Heading to the runway The set of time points that have passed;

[0110] :flight Heading to the runway The set of time points that have passed;

[0111] Step 36: Construct the objective function of the lower-level model, and consider taxiing safety constraints, taxiing sequence constraints, and camera position usage constraints.

[0112] Furthermore, in step 34, the objective function of the upper-level model is to minimize the total runway occupancy time and the total flight delay time, as expressed below:

[0113]

[0114] in, The total runway occupancy time for all aircraft within the planned time period, of which , , and Flights and The time the runway occupancy time; The total delay time for all aircraft within the planned time period, of which , and They represent aircraft Planning and scheduled takeoff and landing times;

[0115] The constraints include:

[0116] Each departing flight can only choose one runway for takeoff:

[0117]

[0118] Occupying the same runway one after another Flight and Runway wake separation constraints between them:

[0119]

[0120] Occupying the same runway one after another Flight and Runway occupancy time constraints between:

[0121]

[0122] Between different runways and Flights approaching each other and Relevant approach slant range constraints:

[0123]

[0124] flight Use of the runway The time cannot exceed its scheduled takeoff / landing time:

[0125]

[0126] Departure flights The delay time cannot exceed its maximum delay time:

[0127]

[0128] track The earliest time slot occupied cannot exceed the earliest time slot in the planned time period:

[0129] .

[0130] Furthermore, in step 36, the objective function of the lower-level model is to minimize the total taxiing time of the aircraft, as expressed below:

[0131]

[0132] in The total taxiing time for all aircraft within the planned time period, of which , and They represent aircraft The runway nodes and parking positions passed through;

[0133] The constraints include:

[0134] Through two consecutive taxiway nodes and Flight It must be greater than the minimum taxiing time for that taxiway segment:

[0135]

[0136] Crossing the taxiway node Flight and It must be greater than the minimum safe taxiing interval:

[0137]

[0138] flight and If passing through the taxiway node Order constraints:

[0139]

[0140] flight and Passing through two consecutive taxiway nodes and Order constraints:

[0141]

[0142] When some aircraft stands are in operation, there are time constraints that prevent other restricted aircraft stands from performing aircraft push-in and push-out operations simultaneously:

[0143]

[0144] Inbound flights The push-in time is affected by the departure flights at the same gate. apron occupancy time constraints:

[0145]

[0146] Departure flights The launch time cannot be earlier than the preceding arriving flight. Minimum transit time constraint:

[0147]

[0148] Departure flights The planned launch time cannot be less than its scheduled launch time:

[0149] .

[0150] Furthermore, step 4 includes the following sub-steps:

[0151] Step 41, for the entire population Initialization is performed, generating 100 individuals. Each individual, representing an aircraft in the taxiing optimization scheme, selects a randomly feasible initial taxiing scheme from the corresponding predicted taxiing path set. This scheme includes runway allocation, runway occupancy time, and time to reach the target runway based on the predicted taxiing path. This allows the calculation of the total runway occupancy time for the upper-level model. Total delay time Total gliding time of the target in the lower-level model At the same time, a maximum penalty coefficient is adopted. Number of skid collisions Runway conflict time Finally, the fitness function is introduced. The overall target value for each individual and whether it meets the constraints are weighted and calculated, where the weights are... , and ;

[0152] Step 42: The upper-level model optimizes runway usage for aircraft through the runway ranking module; for each aircraft's runway selection and runway occupancy time, single-point mutation and single-point crossover operations are performed, using a dynamic mutation probability that gradually decreases with the number of iterations. and dynamic crossover probability By randomly adjusting the runway selection for departing aircraft and fine-tuning the runway usage period for each aircraft, the taxiing time is synchronously adjusted based on the runway start time, thus generating a new generation of individuals for the upper-level model.

[0153] Step 43: After completing the mutation and crossover of the current population, the upper-level model uses the NSGA-II algorithm to... , and To achieve the objective, perform a non-dominated sort on all individuals, select the solution set of the current first frontier, and thus obtain the first Pareto frontier. ;right For each individual in the model, the upper-level model first uses a fitness function based on an elite strategy. Sort the individuals, select the top 50% in fitness, and then, based on crowding distance, select at most the top 50% while maintaining population diversity. Individuals, and finally the optimized population. Pass the data to the lower-level model;

[0154] Step 44: The lower-level model optimizes the aircraft's taxiing time and reduces taxiing conflicts through the taxiing optimization module, using a dynamic mutation probability that gradually decreases with the number of iterations. and dynamic crossover probability The taxiing time of each aircraft is subjected to single-point variation and single-point crossover. By adjusting the taxiing time of each aircraft and resolving existing taxiing conflicts, the runway usage period of the aircraft is adjusted based on the overall change in taxiing time, thereby generating a new generation of individuals for the lower-level model.

[0155] Step 45, after the lower-level model completes the variation and crossover of the current population, , and To achieve the objective, select the best solution at the current first frontier, perform a non-dominated sort on all individuals, select the solution set at the current first frontier, and thus obtain the first Pareto frontier. The lower-level model also uses a fitness function. Calculate and select the top 50% of individuals based on fitness, and select the top 50% based on fitness according to the elite strategy. Each individual, based on crowding distance, selects at most [number] individuals while ensuring population diversity. Each individual will optimize the population. Pass the data to the upper-level model;

[0156] Step 46: After completing one round of upper and lower level model optimization, optimize the current population. Perform individual similarity judgment. If more than 75% of the individuals in the population have high similarity, it indicates that the population may have fallen into a local optimum or converged too early. Select to perform a catastrophe operation, retain the current non-repeating first frontier individual, add 100 random individuals, and return to step 42 to iterate again.

[0157] Step 47: After each round of upper and lower layer model optimization is completed, the model checks whether the current iteration count is not less than the set iteration count. ;like If the algorithm succeeds, it stops and outputs the current optimal solution; otherwise, it stops the iteration count. Return to step 42 to continue the optimization process;

[0158] Step 48: Based on the optimization scheme, evaluate the taxiing optimization results, runway sequencing results, and taxiing hotspot results of the scheme for the performance of the surface aircraft "group" to provide a basis for airport surface operation decisions.

[0159] A trajectory-driven airport surface aircraft taxiing planning system, comprising:

[0160] The airport taxiway trajectory data processing module is used to collect taxiway trajectory data from multiple data sources and perform preprocessing.

[0161] The multi-feature modeling module is used to perform topological modeling based on the physical structure of the airport surface and generate multi-dimensional feature vectors by combining historical taxi trajectory data.

[0162] The temporal trajectory prediction model is used to accurately predict the taxiing trajectory of an individual aircraft and capture the temporal and spatial characteristics of the taxiing trajectory.

[0163] The surface aircraft taxiing performance optimization module is used to combine prediction results and flight plans to establish a two-level optimization model and solve it using the non-dominated sorting multi-objective genetic algorithm (NSGA-II) to optimize taxiing conflicts, delays and runway occupancy.

[0164] Beneficial Effects: This invention provides an aircraft taxiing trajectory prediction and optimization method and system for achieving quantitative analysis and accurate prediction of surface taxiing performance. For individual taxiing behavior, a neural network model is used to capture temporal and spatial characteristics, significantly improving prediction accuracy. For group taxiing behavior, a two-layer optimization model is established by combining prediction results and flight plans, and solved using the NSGA-II algorithm. This provides a novel method for achieving precise surface control and can be widely used in subsequent intelligent surface management applications. It has significant practical application value for improving airport operational efficiency, reducing flight delay rates, and optimizing resource allocation. Attached Figure Description

[0165] Figure 1 This is a flowchart of the method of the present invention;

[0166] Figure 2 This is a topological diagram of the airport surface structure.

[0167] Figure 3 A schematic diagram of the node2vec algorithm for constructing features;

[0168] Figure 4 Diagram of the Bi-LSTM-A model architecture for trajectory prediction;

[0169] Figure 5 A schematic diagram illustrating the calculation of the evaluation index ROUGE-N;

[0170] Figure 6 ROUGE-6 cumulative distribution plot for evaluation results;

[0171] Figure 7 A comparison chart of typical trajectories at the airport;

[0172] Figure 8 A cumulative distribution map of airport surface taxiing time prediction errors;

[0173] Figure 9 Result diagram of airport surface taxiing optimization;

[0174] Figure 10 Gantt chart for runway sorting results of airport surface taxiing optimization;

[0175] Figure 11A comparison chart of taxiing hotspots resulting from airport surface taxiing optimization. Detailed Implementation

[0176] To facilitate understanding of the design objectives, improvements, and combined advantages of this invention, the invention will be further explained in detail with reference to the accompanying drawings and embodiments. It should also be noted that the airport surface structure and taxiway trajectory samples used in this section are merely for illustrating the feasibility of the optimization method and are not exclusively applicable to the processes included in this invention.

[0177] Example 1

[0178] This embodiment selects taxiways from a hub airport as sample data for training and outputs data based on flight schedules, such as... Figure 1 As shown, its operation process is as follows:

[0179] Step 1: Based on the physical structure of the airport surface, perform topological modeling to simplify the complex structure, and introduce the node2vec algorithm from the field of natural language processing to generate surface structure feature vectors; combine historical taxiing trajectory data to extract multi-dimensional feature vectors of aircraft, including taxiing speed, acceleration, latitude and longitude coordinates, etc., to construct a comprehensive feature representation of the aircraft, laying the data foundation for taxiing trajectory prediction. This step includes the following sub-steps:

[0180] (1.1) As Figure 2 As shown, a topological model of the physical structure of the airport surface is performed. By simplifying the complex structure of the airport surface, the layout characteristics of different functional areas within the surface are highlighted, thereby reducing the complexity of the model.

[0181] (1.2) The node2vec node embedding algorithm is used to capture the overall structure of the scene structure graph and the local environment of the sliding nodes, preserve the similarity between nodes, and generate scene structure feature vectors. The specific algorithm design is as follows:

[0182] The graph adjacency matrix constructed based on the scene mixing graph in this invention The `node2vec` algorithm abstracts the entire scene structure into a maximum likelihood optimization problem. Maximum likelihood estimation aims to determine the optimal values ​​of parameters in a probabilistic model using existing observation data, thereby describing the probability and likelihood of the observed data occurring. As a mapping function from taxiway nodes to their corresponding feature representations, For size The matrix is ​​s, where s is the dimension of the feature representation, and for each taxiway node... Then its neighborhood node set Similarly satisfied ,in For the local node set of the taxiway, the subscript is... For taxiway nodes The chosen sampling method aims to maximize the neighborhood log probability density function of all taxiway nodes, as shown in the following expression:

[0183]

[0184] To facilitate optimization and simplify the algorithm, the node feature walking search strategy adopted introduces the following assumptions:

[0185] (1) Neighborhood Independence: For each taxiway node, the probability of exploring its corresponding neighboring node is independent of other neighboring nodes, as shown in the following expression:

[0186]

[0187] (2) Feature symmetry: For each taxiway node, its feature representation as a neighbor of other taxiway nodes is the same as its feature representation in its own space, as shown in the following expression:

[0188]

[0189] in, This indicates the taxiway node. taxiway node The probability of selection The softmax function is used to represent the function for taxiway nodes. The overall probability of selecting all other taxiway nodes.

[0190] Based on these two assumptions, the complex formula can be simplified, and the expression is as follows:

[0191]

[0192] in, The computational cost is too high, so by introducing negative sampling, the multi-classification problem is transformed into multiple binary classification problems for resampling, which greatly reduces the computational load.

[0193] To construct a comprehensive feature set of the scene, this invention employs a random walk strategy to capture the structural relationships and spatial connections between scene nodes. In a random walk, starting from the glide path nodes... Jump to the taxiway node The conditional probability is expressed as follows:

[0194]

[0195] in For the th random walk sequence 1 node Indicates from the taxiway node Jump to the taxiway node The nonnormalized probability is determined by the edge weight. It is a normalization constant obtained through negative sampling.

[0196] like Figure 3 As shown, to establish a comprehensive feature representation of these taxiway nodes, node2vec proposes a random walk strategy using a second-order Markov chain. This means that the selection of the next taxiway node is influenced not only by the current taxiway node but also by the previous one. node2vec controls the bias of the random walk by adjusting two parameters, emphasizing depth-first search (DFS) and breadth-first search (BFS) respectively, thus employing taxiway node walk preference parameters. According to the search strategy, the roaming path just past the glide path node... Jump to the taxiway node From the current point Start moving to the next point The walk, expressed as follows:

[0197]

[0198] in Represents taxiway node and The shortest distance between elements (regardless of edge distance weights). The return parameter represents the probability that the traverse will return to a previously visited gliding path node. This employs a breadth-first search strategy, which is more likely to traverse back to already visited glide path nodes in the next two steps, thus exploring the depth of the starting glide path node's range; if The strategy employed was depth search, which increased the likelihood of navigating to unvisited glide path nodes in the next two steps, thus prompting exploration behavior with a larger range of structures. The in / out parameters represent the breadth of the wander direction. Then a breadth-first search strategy was adopted, which involves walking along already explored structures. The algorithm employs a deep search strategy that walks into unexplored structures.

[0199] (1.3) In order to comprehensively analyze the complex airport operating environment and accurately predict aircraft taxiing dynamics, while taking into account the airport surface taxiing structure, various special structures such as parking stands, runway entrances and exits were also carefully identified. In order to more accurately reflect the constraints when describing the activities of aircraft in the complex airport layout, one-hot coding was used to finely mark the special nodes, thereby improving the prediction accuracy of taxiing behavior of such structures.

[0200] (1.4) Combine historical taxiing trajectory data to extract multi-dimensional feature vectors of aircraft, including information such as taxiing speed, acceleration, latitude and longitude coordinates, and construct a comprehensive feature representation of aircraft.

[0201] (1.5) Combine the scene structure feature vector generated in step (1.3) with the aircraft multidimensional feature vector constructed in step (1.4) to form a complete feature input, providing a data basis for subsequent taxiing trajectory prediction.

[0202] Step two involves constructing a taxiing trajectory temporal prediction model (Bi-LSTM-A) combining a bidirectional long short-term memory network and an attention mechanism, based on the multivariate features generated in step one, for the individual aircraft taxiing trajectory. This model simultaneously captures the temporal relationships and spatial features of the taxiing trajectory. The ROUGE-N evaluation system, root mean square error (RMSE), and mean absolute error (MAE) are used as performance evaluation metrics to verify the model's predictive performance and achieve accurate prediction of the individual aircraft taxiing trajectory. This step includes the following sub-steps:

[0203] (2.1) Conduct in-depth analysis of the individual taxiing trajectories of aircraft in the airport scene, and establish a time series prediction model based on the multi-dimensional characteristics of the aircraft taxiing trajectories that have been constructed.

[0204] (2.2) As Figure 4 As shown, a gliding trajectory temporal prediction model (Bi-LSTM-A) based on a bidirectional long short-term memory network (Bi-LSTM) and an attention mechanism is constructed to simultaneously capture the temporal relationship and spatial features between gliding trajectory points.

[0205] (2.3) The ROUGE-N evaluation system, root mean square error (RMSE), and mean absolute error (MAE) are used as performance evaluation indicators to compare and verify the Bi-LSTM-A model with other traditional models (such as RNN, LSTM, etc.). The specific indicator design is as follows:

[0206] (1) Root Mean Square Error (RMSE)

[0207] RMSE is a commonly used evaluation metric in predictive models. It calculates the square root of the mean square of the differences between predicted and actual values, measuring the discrepancy between the predicted and actual results. The unit of RMSE is the same as that of the original data, making it easier to understand and interpret compared to mean squared error. A smaller RMSE indicates a more accurate prediction and a smaller error. Because RMSE is sensitive to outliers and penalizes larger errors, it is often used to evaluate the accuracy and stability of a model.

[0208]

[0209] (2) Mean Absolute Error (MAE)

[0210] MAE is also a commonly used evaluation metric in forecasting models. It reflects the model's prediction error by calculating the average of the absolute values ​​of the differences between predicted and actual values. It measures the average absolute difference between predicted and true values, as shown in the following expression: The smaller the MAE, the more accurate the model's prediction and the smaller the error. Unlike RMSE, MAE does not square large errors. Therefore, it has low sensitivity to outliers and can only provide a more balanced error assessment.

[0211]

[0212] In the above two formulas, For the first The true value of each sample For the first The predicted value for each sample.

[0213] (3) ROUGE-N evaluation index system (N-gram Co-Occurrence Statistics, ROUGE-N)

[0214] ROUGE-N is one of the ROUGE series metrics. The precision, recall, and F1 score based on ROUGE-N are specifically designed to evaluate n-tuple-level similarity. The ROUGE-N evaluation system can assess the matching rate between predicted and actual gliding trajectories using precision, recall, and F1, as shown in the following expression: Where This indicates that 90% of the n consecutive trajectory nodes (n-points) in the predicted trajectory appear in the sample trajectory, while This indicates that 90% of the n-points in the actual trajectory appear in the predicted trajectory, therefore the result is... It can represent the overall prediction performance of the predicted trajectory compared to the sample trajectory. Therefore, ROUGE-N can accurately measure the proportion of n consecutive nodes that match the generated trajectory with the correct trajectory.

[0215]

[0216]

[0217]

[0218] in This represents the number of n-points in the predicted trajectory that match the sample trajectory. This represents the number of n-points present in the sample trajectory. This represents the number of n-points present in the predicted trajectory.

[0219] like Figure 5 As shown, the ROUGE-N metric value decreases as N increases. Since this embodiment uses a sliding time window of 5, meaning five consecutive trajectory points are used to predict the sixth trajectory point, the six adjacent nodes in the sliding trajectory are directly related. Therefore, this embodiment uses ROUGE-6 and its corresponding precision, recall, and F1 as the model's accuracy evaluation metrics. As a method for evaluating long trajectories, this invention draws on text comparison strategies in the NLP field. Using ROUGE-6 as the trajectory prediction precision has two advantages: by comparing long sequences containing prediction points, it reflects the long-term dependencies of the entire trajectory, thus more comprehensively evaluating the overall model performance, reducing the impact of single-point prediction results on the propagation of overall trajectory errors, and reflecting the continuity and robustness of the prediction model. The results are as follows... Figure 5 As shown.

[0220] (2.4) such as Figure 7 and 8 As shown, the model's prediction accuracy is evaluated through validation analysis of typical taxiing trajectories and taxiing times, and the model outputs accurate predictions of individual aircraft taxiing trajectories.

[0221] Step 3: For the aircraft "group" taxiing performance, based on a precise trajectory prediction dataset, a two-layer spatiotemporal collaborative planning model for surface taxiing is established. The upper layer optimizes runway utilization efficiency, while the lower layer optimizes taxiing operation efficiency. A non-dominated sorting multi-objective genetic algorithm (NSGA-II) is used to optimize and solve the two-layer planning model. An optimized scheme that balances taxiing efficiency and runway operation efficiency is generated through multi-objective optimization algorithms. The optimization results are then verified through simulation and parameter tuning to ensure the feasibility and efficiency of the scheme. This includes the following sub-steps:

[0222] (3.1) Based on the accurate taxiing trajectory prediction dataset and taxiing performance index system, evaluate the taxiing performance under different traffic scenarios to provide data support for optimizing scheduling.

[0223] (3.2) Taking into account the interests of multiple stakeholders such as airports, air traffic control and airlines, a two-layer planning model for surface taxiing based on accurate trajectory prediction is established from the perspective of integrated scheduling of the arrival and departure taxiing process.

[0224] (3.3) The relevant sets, parameters, and variables required by the upper-level model are as follows:

[0225] variable:

[0226] If the flight Prior to flights Start occupying the runway but ,otherwise ;

[0227] If the flight The runway selection is but ,otherwise ;

[0228] :flight On the runway Planned take-off and landing times;

[0229] parameter:

[0230] Number is Flights;

[0231] Approaching flights On the flight The wake spacing ahead;

[0232] :track and The relevant approach slant ranges between them;

[0233] :flight The final approach velocity;

[0234] :flight The time the runway occupancy time;

[0235] :flight On the runway The planned takeoff and landing times;

[0236] The earliest time in the current planning period;

[0237] :track Start time of use;

[0238] :track End time of use;

[0239] :track Conflict time of use;

[0240] :flight The maximum acceptable delay time;

[0241] gather:

[0242] : The scheduled time for the assembly of flights, among which ,at the same time This indicates that there are a total of [number] people on the field within the planned time. Flights;

[0243] Arrival of flights within the planned timeframe;

[0244] Assembly point for departing flights within the planned timeframe;

[0245] The set of runways to choose from, among which ;

[0246] Airport surface taxiway node collection, among which ;

[0247] The set of nodes of the airport surface runway;

[0248] Airport surface parking stand node collection;

[0249] :flight Heading to the runway The set of nodes for the predicted gliding path;

[0250] :flight Heading to the runway The set of nodes for the shortest glide path;

[0251] :flight Heading to the runway The set of time points that have passed;

[0252] :flight Heading to the runway The set of time points that have passed.

[0253] (3.4) The objective function used in the construction of the upper-level model is shown below, and it takes into account runway selection constraints, runway spacing constraints, runway occupancy constraints, optimization range constraints, and maximum delay constraints:

[0254] Objective function:

[0255] Runways are a critical resource for airport operations. The upper-level optimization model primarily considers the needs of air traffic control and airlines, aiming to minimize total runway occupancy time while maximizing runway utilization efficiency through rational resource scheduling and allocation. This optimization process needs to consider not only flight takeoff and landing sequences but also flight delay times. The objective function is expressed as:

[0256]

[0257] in The total runway occupancy time for all aircraft within the planning period is used to assess runway utilization efficiency by measuring runway occupancy time during the planning period. , , and Flights and The time the runway occupancy time; The total delay time for all aircraft within the planned time period is used to measure the overall flight delay situation during the planned time period. , and They represent aircraft The planning and scheduled take-off and landing times.

[0258] Constraints:

[0259] Each departing flight can only choose one runway for takeoff:

[0260]

[0261] Occupying the same runway one after another Flight and Runway wake separation constraints between them:

[0262]

[0263] Occupying the same runway one after another Flight and Runway occupancy time constraints between:

[0264]

[0265] Between different runways and Flights approaching each other and Relevant approach slant range constraints:

[0266]

[0267] flight Use of the runway The time cannot exceed its scheduled takeoff / landing time:

[0268]

[0269] Departure flights The delay time cannot exceed its maximum delay time:

[0270]

[0271] track The earliest time slot occupied cannot exceed the earliest time slot in the planned time period:

[0272]

[0273] (3.5) The relevant sets, parameters, and variables required by the lower-level model are shown below:

[0274] variable:

[0275] :flight go through The node's time, if not elapsed ;

[0276] If the flight The path passes through the taxiway node. but ,otherwise ;

[0277] If the flight Prior to flights Crossing the taxiway node but ,otherwise ;

[0278] parameter:

[0279] Number is taxiway nodes;

[0280] :flight and Minimum safe taxiing distance between them;

[0281] taxiway node and The minimum distance between them;

[0282] :flight Passing through the taxiway node speed;

[0283] Minimum turnaround time for flights;

[0284] : Apron occupancy time for flights;

[0285] Departure flights The planned launch time;

[0286] Number of skiing incidents during the planned time period;

[0287] : Maximum constant value;

[0288] gather:

[0289] : The scheduled time for the assembly of flights, among which ,at the same time This indicates that there are a total of [number] people on the field within the planned time. Flights;

[0290] Arrival of flights within the planned timeframe;

[0291] Assembly point for departing flights within the planned timeframe;

[0292] The set of runways to choose from, among which ;

[0293] Airport surface taxiway node collection, among which ;

[0294] The set of nodes of the airport surface runway;

[0295] Airport surface parking stand node collection;

[0296] :flight Heading to the runway The set of nodes for the predicted gliding path;

[0297] :flight Heading to the runway The set of nodes for the shortest glide path;

[0298] :flight Heading to the runway The set of time points that have passed;

[0299] :flight Heading to the runway The set of time points that have passed.

[0300] (3.6) The objective function used in the construction of the lower-level model is shown below, and taxiing safety constraints, taxiing sequence constraints, and camera position usage constraints are taken into account:

[0301] Objective function:

[0302] Taxiways are a critical resource for airport surface operations. The lower-level optimization model primarily considers the needs of the airport and airlines, aiming to minimize the overall taxiing time of aircraft, as shown in the following expression:

[0303]

[0304] in The total taxiing time for all aircraft within the planned time period, of which , and They represent aircraft The runway nodes and parking positions passed through.

[0305] Constraints:

[0306] Through two consecutive taxiway nodes and Flight It must be greater than the minimum taxiing time for that taxiway segment:

[0307]

[0308] Crossing the taxiway node Flight and It must be greater than the minimum safe taxiing interval:

[0309]

[0310] flight and If passing through the taxiway node Order constraints:

[0311]

[0312] flight and Passing through two consecutive taxiway nodes and Order constraints:

[0313]

[0314] When some aircraft stands are in operation, there are time constraints that prevent other restricted aircraft stands from performing aircraft push-in and push-out operations simultaneously:

[0315]

[0316] Inbound flights The push-in time is affected by the departure flights at the same gate. apron occupancy time constraints:

[0317]

[0318] Departure flights The launch time cannot be earlier than the preceding arriving flight. Minimum transit time constraint:

[0319]

[0320] Departure flights The planned launch time cannot be less than its scheduled launch time:

[0321]

[0322] Step four involves solving the bi-level programming model using the NSGA-II algorithm to obtain the surface optimization scheme. The optimization results are then verified using examples from typical domestic airports, and compared with traditional taxiing optimization schemes to validate the model's efficiency and practicality. This step includes the following sub-steps:

[0323] (4.1) The model first applies to the entire population. Initialization is performed, generating 100 individuals. Each individual, representing an aircraft in the taxiing optimization scheme, selects a randomly feasible initial taxiing scheme from the corresponding predicted taxiing path set. This scheme includes runway allocation, runway occupancy time, and time to reach the target runway based on the predicted taxiing path. This allows the calculation of the total runway occupancy time for the upper-level model. Total delay time Total gliding time of the target in the lower-level model At the same time, a maximum penalty coefficient is adopted. Number of skid collisions Runway conflict time Finally, the fitness function is introduced. The overall target value for each individual and whether it meets the constraints are weighted and calculated, where the weights are... , and .

[0324] (4.2) The upper-level model incorporates a runway ranking module to optimize runway utilization for aircraft. This module performs single-point mutation and single-point crossover operations for each aircraft's runway selection and runway occupancy time. To enhance convergence in long-term iterations, a dynamic mutation probability that gradually decreases with the number of iterations is employed. and dynamic crossover probability By randomly adjusting the runway selection for departing aircraft and fine-tuning the runway usage period for each aircraft, the taxiing time can be synchronously adjusted based on the runway start time, thus generating a new generation of individuals for the upper-level model.

[0325] (4.3) After completing the mutation and crossover of the current population, the upper-level model uses the NSGA-II algorithm to... , and To achieve the objective, perform a non-dominated sort on all individuals, select the solution set of the current first frontier, and thus obtain the first Pareto frontier. .right For each individual in the model, the upper-level model first uses a fitness function based on an elite strategy. Sort the individuals, select the top 50% in fitness, and then, based on crowding distance, select at most the top 50% while maintaining population diversity. Individuals, and finally the optimized population. Pass it to the lower-level model.

[0326] (4.4) In the lower-level model, a taxiing optimization module was designed to optimize the aircraft's taxiing over-point time and reduce taxiing conflicts, using a dynamic mutation probability that gradually decreases with the number of iterations. and dynamic crossover probability The taxiing time of each aircraft is subjected to single-point variation and single-point crossover. By adjusting the taxiing time of each aircraft and resolving existing taxiing conflicts, the runway usage periods of the aircraft are adjusted based on the overall change in taxiing time, thereby generating a new generation of individuals for the lower-level model.

[0327] (4.5) The lower-level model, after completing the variation and crossover of the contemporary population, uses , and To achieve the objective, select the best solution at the current first frontier, perform a non-dominated sort on all individuals, select the solution set at the current first frontier, and thus obtain the first Pareto frontier. The lower-level model also uses a fitness function. Calculate and select the top 50% of individuals based on fitness, and select the top 50% based on fitness according to the elite strategy. Each individual, based on crowding distance, selects at most [number] individuals while ensuring population diversity. Each individual will optimize the population. Pass it to the upper-level model.

[0328] (4.6) After completing one round of upper and lower level model optimization, optimize the current population. Perform individual similarity judgment. If more than 75% of the individuals in the population have high similarity, it indicates that the population may have fallen into a local optimum or converged too early. Select to perform a catastrophe operation, retain the current non-repeating first frontier individual, add 100 random individuals, and return to step 2 to iterate again.

[0329] (4.7) After each round of optimization of the upper and lower layer models, the model checks whether the current iteration number is not less than the set iteration number. .like If the algorithm succeeds, it stops and outputs the current optimal solution; otherwise, it stops the iteration count. Return to step 2 to continue the optimization process.

[0330] (4.8) Based on the optimized scheme, evaluate the taxiing optimization results, runway sequencing results, and taxiing hotspot results for the performance of the surface aircraft "group," such as... Figure 9-11 As shown, this can provide a basis for airport surface operation decisions.

[0331] Example 2

[0332] This embodiment provides a trajectory-driven airport surface aircraft taxiing planning system for implementing the method of Embodiment 1. The system includes the following modules:

[0333] (a) Airport taxiway trajectory data processing module

[0334] This module is used to collect and preprocess gliding trajectory data from multiple data sources.

[0335] (II) Multi-feature construction modeling module

[0336] This module is used for topology modeling based on the physical structure of the airport surface, and generates multi-dimensional feature vectors by combining historical taxiway trajectory data, providing a data foundation for taxiway trajectory prediction.

[0337] Specifically, this module performs the following operations: It performs topological modeling of the airport surface's physical structure, simplifying complex structures and highlighting the layout features of different functional areas within the surface to reduce model complexity; it employs the node2vec embedding algorithm to capture the overall structure of the surface structure diagram and the local environment of taxiing nodes, preserving the similarity between nodes to generate surface structure feature vectors; it uses one-hot encoding to finely label special nodes, including parking stands, runway entrances, and exits, thereby improving the accuracy of taxiing behavior prediction for such structures; it combines historical taxiing trajectory data to extract multi-dimensional feature vectors of aircraft, including taxiing speed, acceleration, latitude and longitude coordinates, to construct a comprehensive feature representation of the aircraft; and it combines the generated surface structure feature vectors with the constructed multi-dimensional aircraft feature vectors to form a complete feature input, providing a data foundation for subsequent taxiing trajectory prediction.

[0338] (III) Time-series trajectory prediction model

[0339] This model is used to accurately predict the taxiing trajectories of individual aircraft, capturing the temporal and spatial characteristics of the taxiing trajectories.

[0340] Specifically, the model performs the following operations: It conducts in-depth analysis of the individual taxiing trajectories of aircraft in the airport environment, establishing a temporal prediction model based on the multi-dimensional features of the aircraft taxiing trajectories; it constructs a taxiing trajectory temporal prediction model (Bi-LSTM-A) based on a bidirectional long short-term memory network (Bi-LSTM) and an attention mechanism to simultaneously capture the temporal relationships and spatial features between taxiing trajectory points; it uses the ROUGE-N evaluation system, root mean square error (RMSE), and mean absolute error (MAE) as performance evaluation indicators to compare and validate the Bi-LSTM-A model with other traditional models; and it evaluates the model's prediction accuracy through validation analysis of typical taxiing trajectories and taxiing times, outputting accurately predicted individual aircraft taxiing trajectories.

[0341] (iv) Surface aircraft taxiing performance optimization module

[0342] This module is used to combine prediction results and flight plans to establish a two-level optimization model and solve it using the NSGA-II algorithm, thereby optimizing taxiing conflicts, delays and runway occupancy, and improving group taxiing efficiency and surface operation safety.

[0343] Specifically, this module performs the following operations: Based on a precise taxiing trajectory prediction dataset and taxiing performance index system, it evaluates taxiing performance under different traffic flow scenarios, providing data support for optimized scheduling; considering the interests of multiple stakeholders such as airports, air traffic control, and airlines, it establishes a two-tiered spatiotemporal collaborative planning model for airport surface taxiing based on precise trajectory prediction, from the perspective of integrated scheduling of arrival and departure taxiing processes; it optimizes the two-tiered planning model using a non-dominated sorting multi-objective genetic algorithm (NSGA-II), generating an optimized scheme that balances taxiing efficiency and runway operating efficiency through multi-objective optimization algorithms, and performing simulation verification and parameter tuning on the optimization results to ensure the feasibility and efficiency of the scheme; through case studies of typical domestic airports, it compares the optimization results with traditional taxiing optimization schemes to verify the efficiency and practicality of the model; and it applies the optimized scheme to airport surface operation decisions, providing scientific support for the coordinated improvement of taxiing and runway operating efficiency, and realizing intelligent and precise control of airport surface operations.

[0344] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A trajectory-driven airport surface aircraft taxiing planning method, characterized in that: Includes the following steps: Step 1: Perform topological modeling based on the physical structure of the airport surface and introduce the node2vec algorithm to generate surface structure feature vectors; combine historical taxiing trajectory data to extract multi-dimensional feature vectors of aircraft and construct a comprehensive feature representation of the aircraft. Step 2: For the taxiing trajectory of an aircraft "individual", based on the multi-dimensional features generated in Step 1, construct a taxiing trajectory temporal prediction model that combines a bidirectional long short-term memory network with an attention mechanism to achieve accurate prediction of the taxiing trajectory of an aircraft "individual". Step 3: For the taxiing performance of aircraft "groups", based on the accurate trajectory prediction dataset, establish a two-layer planning model for surface taxiing spatiotemporal coordination, in which the upper-layer model optimizes runway utilization efficiency and the lower-layer model optimizes taxiing operation efficiency. Step 4: The non-dominated sorting multi-objective genetic algorithm is used to optimize and solve the bi-level programming model, generating an optimized scheme that takes into account both taxiing efficiency and runway operation efficiency.

2. The method according to claim 1, characterized in that: Step 1 includes the following sub-steps: Step 11: Perform topological modeling of the physical structure of the airport surface. By simplifying the complex structure of the airport surface and highlighting the layout characteristics of different functional areas within the surface, the complexity of the model can be reduced. Step 12: The node2vec node embedding algorithm is used to capture the overall structure of the scene structure graph and the local environment of the sliding nodes, preserve the similarity between nodes, and generate scene structure feature vectors. Step 13: Use one-hot coding to finely mark special nodes, including parking positions, runway entrances and exits, thereby improving the accuracy of taxiing behavior prediction for such structures; Step 14: Combine historical taxiing trajectory data to extract multi-dimensional feature vectors of the aircraft, including taxiing speed, acceleration, and latitude and longitude coordinates, and construct a comprehensive feature representation of the aircraft. Step 15: Combine the scene structure feature vector generated in Step 13 with the aircraft multidimensional feature vector constructed in Step 14 to form a complete feature input, providing a data foundation for subsequent taxiing trajectory prediction.

3. The method according to claim 2, characterized in that, Step 12 is as follows: Graph adjacency matrix constructed from the scene mixing graph The entire scene structure is abstracted into a maximum likelihood optimization problem using node2vec; where As a mapping function from taxiway nodes to their corresponding feature representations, For size The matrix is ​​s, where s is the dimension of the feature representation, and for each taxiway node... Then its neighborhood node set Similarly satisfied ,in For the local node set of the taxiway, the subscript is... For taxiway nodes The chosen sampling method aims to maximize the neighborhood log probability density function of all taxiway nodes, as shown in the following expression: To facilitate optimization, the following assumptions are introduced: (1) Neighborhood Independence: For each taxiway node, the probability of exploring its corresponding neighboring node is independent of other neighboring nodes, as shown in the following expression: (2) Feature symmetry: For each taxiway node, its feature representation as a neighbor of other taxiway nodes is the same as its feature representation in its own space, as shown in the following expression: in, This indicates the taxiway node. taxiway node The probability of selection The softmax function is used to represent the function for taxiway nodes. The overall probability of selecting all other taxiway nodes; Based on the above assumptions, the complex formula is simplified as follows: in, Because of the excessively high computational cost, the multi-classification problem is transformed into multiple binary classification problems for resampling by introducing negative sampling. A random walk strategy is employed to capture the structural relationships and spatial connections between scene nodes, starting from the glide path nodes within a random walk. Jump to the taxiway node The conditional probability is expressed as follows: in For the th random walk sequence 1 node Indicates from the taxiway node Jump to the taxiway node The nonnormalized probability is determined by the edge weight. It is a normalization constant obtained through negative sampling; A random walk strategy using a second-order Markov chain is employed, where the selection of the next taxiway node is influenced by both the current and previous taxiway nodes. Two parameters are adjusted to control the bias of the random walk, emphasizing depth-first search and breadth-first search respectively, thus employing taxiway node walk preference parameters. According to the search strategy, the roaming path just past the glide path node... Jump to the taxiway node From the current point Start moving to the next point The walk, expressed as follows: in Represents taxiway node and The shortest distance between elements is independent of the distance weight of the edges; The return parameter represents the probability that the traverse will return to a previously visited gliding path node. If so, a breadth-first search strategy is adopted; if If so, a depth-first search strategy is adopted; The in / out parameters represent the breadth of the wander direction. Then a breadth-first search strategy is adopted, which walks through already explored structures. Then, a depth search strategy is adopted to walk into unexplored structures.

4. The method according to claim 1, characterized in that: Step 2 includes the following sub-steps: Step 21: Conduct in-depth analysis of the individual taxiing trajectories of aircraft in the airport area, and establish a time series prediction model based on the constructed multi-dimensional characteristics of aircraft taxiing trajectories; Step 22: Construct a gliding trajectory temporal prediction model based on a bidirectional long short-term memory network and an attention mechanism to simultaneously capture the temporal relationship and spatial features between gliding trajectory points; Step 23: The ROUGE-N evaluation system, root mean square error and mean absolute error are used as performance evaluation indicators to evaluate the prediction accuracy of the gliding trajectory time series prediction model. Step 24: Through verification analysis of typical taxiing trajectories and taxiing times, evaluate the prediction accuracy of the model and output the accurate predicted taxiing trajectory of the aircraft "individual".

5. The method according to claim 4, characterized in that: The specific indicator design for step 23 is as follows: (1) Root mean square error (2) Mean Absolute Error In the above two formulas, For the first The true value of each sample For the first Predicted values ​​for each sample; (3) ROUGE-N evaluation index system in This represents the number of n-points in the predicted trajectory that match the sample trajectory. This represents the number of n-points present in the sample trajectory. This represents the number of n-points present in the predicted trajectory.

6. The method according to claim 1, characterized in that: Step 3 includes the following sub-steps: Step 31: Based on the accurate taxiing trajectory prediction dataset and taxiing performance index system, evaluate taxiing performance under different traffic scenarios to provide data support for optimizing scheduling; Step 32: Establish a two-layer planning model for spatiotemporal coordination of surface gliding based on accurate trajectory prediction, including an upper-layer model and a lower-layer model; Step 33: The relevant sets, parameters, and variables required by the upper-level model are shown below: variable: If the flight Prior to flights Start occupying the runway but ,otherwise ; If the flight The runway selection is but ,otherwise ; :flight On the runway The planned takeoff and landing times; parameter: Number is Flights; Approaching flights On the flight The wake spacing ahead; :track and The relevant approach slant ranges between them; :flight The final approach velocity; :flight The time the runway occupies; :flight On the runway The planned takeoff and landing times; The earliest time in the current planning period; :track Start time of use; :track End time of use; :track Conflict time of use; :flight The maximum acceptable delay time; gather: : The scheduled time for the assembly of flights, among which ,at the same time This indicates that there are a total of [number] people on the field within the planned time. Flights; : Arrange for flights to arrive within the planned timeframe; Assembly point for departing flights within the planned timeframe; The set of runways to choose from, among which ; Airport surface taxiway node collection, among which ; The set of nodes of the airport surface runway; Airport surface parking stand node collection; :flight Heading to the runway The set of nodes for the predicted gliding path; :flight Heading to the runway The set of nodes for the shortest glide path; :flight Heading to the runway The set of time points that have passed; :flight Heading to the runway The set of time points that have passed; Step 34: Construct the objective function of the upper-level model, and consider runway selection constraints, runway spacing constraints, runway occupancy constraints, optimization range constraints, and maximum delay constraints. Step 35, the relevant sets, parameters and variables required by the lower-level model are as follows: variable: :flight go through The node's time, if not elapsed ; If the flight The path passes through the taxiway node. but ,otherwise ; If the flight Prior to flights Crossing the taxiway node but ,otherwise ; parameter: Number is taxiway nodes; :flight and Minimum safe taxiing distance between them; Taxiway node and The minimum distance between them; :flight Passing through the taxiway node speed; Minimum turnaround time for flights; : Apron occupancy time for flights; Departure flights The planned launch time; Number of skiing incidents during the planned time period; : Maximum constant value; gather: : The scheduled time for the assembly of flights, among which ,at the same time This indicates that there are a total of [number] people on the field within the planned time. Flights; : Arrange for flights to arrive within the planned timeframe; Assembly point for departing flights within the planned timeframe; The set of runways to choose from, among which ; Airport surface taxiway node collection, among which ; The set of nodes of the airport surface runway; Airport surface parking stand node collection; :flight Heading to the runway The set of nodes for the predicted gliding path; :flight Heading to the runway The set of nodes for the shortest glide path; :flight Heading to the runway The set of time points that have passed; :flight Heading to the runway The set of time points that have passed; Step 36: Construct the objective function of the lower-level model, and consider taxiing safety constraints, taxiing sequence constraints, and camera position usage constraints.

7. The method according to claim 6, characterized in that: In step 34, the objective function of the upper-level model is to minimize the total runway occupancy time and the total flight delay time, as expressed below: in, The total runway occupancy time for all aircraft within the planned time period, of which , , and Flights and The time the runway occupies; The total delay time for all aircraft within the planned time period, of which , and They represent aircraft Planning and scheduled takeoff and landing times; The constraints include: Each departing flight can only choose one runway for takeoff: Occupying the same runway one after another Flight and Runway wake separation constraints between them: Occupying the same runway one after another Flight and Runway occupancy time constraints between: Between different runways and Flights approaching each other and Relevant approach slant range constraints: flight Use of the runway The time cannot exceed its scheduled takeoff / landing time: Departure flights The delay time cannot exceed its maximum delay time: track The earliest time slot occupied cannot exceed the earliest time slot in the planned time period: 。 8. The method according to claim 6, characterized in that: In step 36, the objective function of the lower-level model is to minimize the total taxiing time of the aircraft, as expressed below: in The total taxiing time for all aircraft within the planned time period, of which , and They represent aircraft The runway nodes and parking positions passed through; The constraints include: Through two consecutive taxiway nodes and Flight It must be greater than the minimum taxiing time for that taxiway segment: Crossing the taxiway node Flight and It must be greater than the minimum safe taxiing interval: flight and If passing through the taxiway node Order constraints: flight and Passing through two consecutive taxiway nodes and Order constraints: When some aircraft stands are in operation, there are time constraints that prevent other restricted aircraft stands from performing aircraft push-in and push-out operations simultaneously: Inbound flights The push-in time is affected by the departure flights at the same gate. apron occupancy time constraints: Departure flights The launch time cannot be earlier than the preceding arriving flight. Minimum transit time constraint: Departure flights The planned launch time cannot be less than its scheduled launch time: 。 9. The method according to claim 1, characterized in that: Step 4 includes the following sub-steps: Step 41, for the entire population Initialization is performed, generating 100 individuals. Each individual, representing an aircraft in the taxiing optimization scheme, selects a randomly feasible initial taxiing scheme from the corresponding predicted taxiing path set. This scheme includes runway allocation, runway occupancy time, and time to reach the target runway based on the predicted taxiing path. This allows the calculation of the total runway occupancy time for the upper-level model. Total delay time Total gliding time of the target in the lower-level model At the same time, a maximum penalty coefficient is adopted. Number of skid collisions Runway conflict time Finally, the fitness function is introduced. The overall target value for each individual and whether it meets the constraints are weighted and calculated, where the weights are... , and ; Step 42: The upper-level model optimizes runway usage for aircraft through the runway ranking module; for each aircraft's runway selection and runway occupancy time, single-point mutation and single-point crossover operations are performed, using a dynamic mutation probability that gradually decreases with the number of iterations. and dynamic crossover probability By randomly adjusting the runway selection for departing aircraft and fine-tuning the runway usage period for each aircraft, the taxiing time is synchronously adjusted based on the runway start time, thus generating a new generation of individuals for the upper-level model. Step 43: After completing the mutation and crossover of the current population, the upper-level model uses the NSGA-II algorithm to... , and To achieve the objective, perform a non-dominated sort on all individuals, select the solution set of the current first frontier, and thus obtain the first Pareto frontier. ;right For each individual in the model, the upper-level model first uses a fitness function based on an elite strategy. Sort the individuals, select the top 50% in fitness, and then, based on crowding distance, select at most the top 50% while maintaining population diversity. Individuals, and finally the optimized population Pass the data to the lower-level model; Step 44: The lower-level model optimizes the aircraft's taxiing time and reduces taxiing conflicts through the taxiing optimization module, using a dynamic mutation probability that gradually decreases with the number of iterations. and dynamic crossover probability The taxiing time of each aircraft is subjected to single-point variation and single-point crossover. By adjusting the taxiing time of each aircraft and resolving existing taxiing conflicts, the runway usage period of the aircraft is adjusted based on the overall change in taxiing time, thereby generating a new generation of individuals for the lower-level model. Step 45, after the lower-level model completes the variation and crossover of the current population, , and To achieve the objective, select the best solution at the current first frontier, perform a non-dominated sort on all individuals, select the solution set at the current first frontier, and thus obtain the first Pareto frontier. The lower-level model also uses a fitness function. Calculate and select the top 50% of individuals based on fitness, and select the top 50% based on fitness according to the elite strategy. Each individual, based on crowding distance, selects at most [number] individuals while ensuring population diversity. Each individual will optimize the population. Pass the data to the upper-level model; Step 46: After completing one round of upper and lower level model optimization, optimize the current population. Perform individual similarity judgment. If more than 75% of the individuals in the population have high similarity, it indicates that the population may have fallen into a local optimum or converged too early. Select to perform a catastrophe operation, retain the current non-repeating first frontier individual, add 100 random individuals, and return to step 42 to iterate again. Step 47: After each round of upper and lower layer model optimization is completed, the model checks whether the current iteration count is not less than the set iteration count. ;like If the algorithm succeeds, it stops and outputs the current optimal solution; otherwise, it stops the iteration count. Return to step 42 to continue the optimization process; Step 48: Based on the optimization scheme, evaluate the taxiing optimization results, runway sequencing results, and taxiing hotspot results of the scheme for the performance of the surface aircraft "group" to provide a basis for airport surface operation decisions.

10. A trajectory-driven airport surface aircraft taxiing planning system, used to implement the method of claim 1, characterized in that: include: The airport taxiway trajectory data processing module is used to collect taxiway trajectory data from multiple data sources and perform preprocessing. The multi-feature modeling module is used to perform topological modeling based on the physical structure of the airport surface and generate multi-dimensional feature vectors by combining historical taxi trajectory data. The temporal trajectory prediction model is used to accurately predict the taxiing trajectory of an individual aircraft and capture the temporal and spatial characteristics of the taxiing trajectory. The surface aircraft taxiing performance optimization module is used to combine prediction results and flight plans to establish a two-level optimization model and solve it using a non-dominated sorting multi-objective genetic algorithm to optimize taxiing conflicts, delays and runway occupancy.

Images