Airport operation service resource dynamic allocation and scheduling system and method

Abstract

This invention belongs to the field of data processing technology, specifically relating to a dynamic allocation and scheduling system and method for airport operation service resources. It includes: a relationship graph construction module for generating an airport resource-flight service relationship graph; a graph attention network processing module for performing multi-layer, multi-head graph attention aggregation on the airport resource-flight service relationship graph and decoding resource embedding vectors and task embedding vectors into non-negative classification scheduling indices; a cellular automata evolution module for generating candidate resource allocation scenarios; and a Stackelberg game scheduling module for generating a final resource allocation and scheduling scheme through conflict detection and game iteration, and issuing resource scheduling execution instructions to the execution terminal. This invention solves the problems in existing technologies such as the lack of unified representation in independent modeling of various resources, the lack of differentiation between fixed and movable resources, and the lack of game coordination mechanisms among multiple decision-making entities.

Description

Technical Field

[0001] This invention belongs to the field of data processing technology, specifically relating to a dynamic allocation and scheduling system and method for airport operation service resources. Background Technology

[0002] With the continuous growth of civil aviation transport volume, large and medium-sized hub airports are simultaneously handling an increasing number of flights and involving a growing variety of service resources. The coordinated scheduling of multiple service resources, such as boarding gates, jet bridges, shuttle buses, ground staff, and baggage sorting channels, has become a key factor affecting the overall operational efficiency of airports. Currently, the allocation and scheduling of airport operational service resources mainly relies on a combination of manual experience-based scheduling and rule-based information systems. Scheduling personnel assign service resources to each flight based on flight schedules and established priority rules. This method can basically meet the needs when the number of flights is small and the operating environment is stable. However, when flight density increases, flight delays occur frequently, or multiple flights arrive simultaneously within a short period, manual scheduling struggles to comprehensively consider the spatiotemporal constraints of all resource categories and the competitive relationships between multiple airlines within a limited timeframe. This can easily lead to an imbalance where local resource overload coexists with resource idleness in other areas. In academic and engineering fields, existing research has applied methods such as integer programming, genetic algorithms, or ant colony optimization to optimize single aspects such as boarding gate allocation or ground vehicle routing planning, achieving some positive results. However, these methods typically model and solve various resources separately, lacking a unified representation of the complex relationships between resources and flights, and failing to capture the coupling constraints between different types of resources. Furthermore, most existing methods assume complete transparency of resource supply and demand information globally, ignoring the strategic game characteristics arising from multiple decision-makers pursuing their own self-interest in actual airport operations. This results in insufficient applicability of the generated scheduling schemes in multi-agent coordination scenarios. In addition, existing spatial resource situational analysis methods lack differentiation and processing of the fundamental differences in spatial allocation mechanisms between fixed and mobile resources. They treat immovable resources such as boarding gates and jet bridges as a single entity for unified scheduling with mobile resources such as shuttle buses and ground staff, generating schemes inconsistent with the airport's physical operating mechanisms and difficult to implement in practice. Summary of the Invention

[0003] Therefore, the main objective of this invention is to provide a dynamic allocation and scheduling system and method for airport operation service resources, which solves the problems in the prior art of lacking unified representation for independent modeling of various resources, failing to distinguish between fixed and mobile resources, and lacking a game coordination mechanism among multiple decision-making entities.

[0004] The technical solution adopted in this invention is as follows: The airport operation service resource dynamic allocation and scheduling system includes: a relationship graph construction module, which uses available airport service resources as resource nodes and flights to be supported as task nodes, establishes connections between resource nodes and task nodes based on service matching conditions, and generates an airport resource-flight service relationship graph; a graph attention network processing module, which performs multi-layer multi-head graph attention aggregation on the airport resource-flight service relationship graph, outputs the resource embedding vector of each resource node and the task embedding vector of each task node, and decodes the resource embedding vector and task embedding vector into non-negative classification scheduling indices; and a cellular automaton algorithm. The first module discretizes the airport's planar area into a cellular grid, injects categorized scheduling indicators into each cell, and through synchronous iterative evolution, diffuses the scheduling indicators of movable category resources among adjacent cells towards spatial equilibrium, generating a resource candidate allocation situation. The second module, a Stackelberg game scheduling module, uses the airport management as the leader and each flight support task agent as the follower. The leader generates and iteratively updates a priority scheduling strategy based on the resource candidate allocation situation, and the followers select service resources according to the priority scheduling strategy. Through conflict detection and game iteration, the final resource allocation scheduling scheme is generated and resource scheduling execution instructions are issued to the execution terminal.

[0005] Furthermore, service resources are divided into 5 categories: Category 1 is boarding gates, Category 2 is jet bridges, Category 3 is shuttle buses, Category 4 is ground staff groups, and Category 5 is baggage sorting channels; Categories 1, 2, and 5 are fixed resources, and Categories 3 and 4 are mobile resources.

[0006] Furthermore, for each resource node, the resource's geographical coordinates, current occupancy status identifier, remaining time until the next release, and category number are collected and arranged in the collection order to form the initial feature vector of the resource node; for each task node, the geographical coordinates of the parking position expected to be used by the flight, the number of seats corresponding to the flight type, and the expected arrival or departure time are collected and arranged in the collection order to form the initial feature vector of the task node; when a resource node and a task node simultaneously satisfy aircraft type adaptation, service type matching, and the overlap duration of the resource availability time window and the flight guarantee time window is not less than the minimum continuous operation duration of the corresponding category of service resources, an undirected edge is established between the two.

[0007] Furthermore, the graph attention network comprises L consecutively connected graph attention layers, each containing K independent attention heads. Within each layer, for each attention head, the following operations are performed on each node in the graph: traverse all neighboring nodes directly connected to the current node via undirected edges; concatenate the current node's feature vector with the feature vectors of each neighboring node, and feed them into the corresponding single-layer feedforward neural network of the attention head. After linear mapping and LeakyReLU activation, a scalar is output as the original attention score of the corresponding neighboring node; perform softmax normalization on the original attention scores of all neighboring nodes to obtain the normalized attention coefficients for each neighboring node; multiply the feature vector of each neighboring node by the corresponding normalized attention coefficient and sum element-wise to obtain a single-head aggregation vector; concatenate the K single-head aggregation vectors to form the output feature vector of the current node in the current layer and pass it to the next layer as input; after processing by L layers, output the resource embedding vectors of each resource node and the task embedding vectors of each task node.

[0008] Furthermore, each resource embedding vector is input into the pre-trained resource decoding fully connected layer, and after ReLU activation, it outputs a non-negative scalar as the resource schedulability of the corresponding resource node; each task embedding vector is input into the pre-trained task decoding fully connected layer, and after ReLU activation, it outputs 5 non-negative scalars, which correspond to the single-class task demand intensity of the 5 categories in turn.

[0009] Furthermore, the cellular automaton evolution module divides the plane area of ​​the airport apron and terminal building into equally spaced square grids, with each grid cell serving as a cell. For each cell, it retrieves all resource nodes and all task nodes whose geographical coordinates fall within the cell's coverage area. For each of the five categories, it sums the resource schedulability of all resource nodes within the cell's coverage area that belong to the corresponding category as the cell's category resource index value, and sums the single-category task demand intensity of all task nodes within the cell's coverage area that belong to the corresponding category as the cell's category task index value.

[0010] Furthermore, at the start of each evolution step, a global snapshot backup of the current category index values ​​of all cells is performed. Based on the snapshot state, diffusion operations are performed independently only for movable category resources: For each movable category, the snapshot category resource index values ​​and snapshot category task index values ​​of the eight neighboring cells within the current cell's neighborhood are read. Neighboring cells whose category resource index values ​​are strictly greater than their category task index values ​​are marked as surplus cells, and those whose category resource index values ​​are strictly less than their category task index values ​​are marked as scarce cells. For each surplus cell, its category resource index value in the corresponding movable category is equally split. The process is divided into 9 parts, with 1 part reserved for the surplus cell itself, and the remaining 8 parts allocated to the scarce cells in descending order of their category task index values ​​in the corresponding mobile category. If the number of scarce cells in the corresponding mobile category within the Moore neighborhood is 0, the category resource index value of the surplus cell in the corresponding mobile category remains unchanged. The category resource index value of the fixed category resource remains unchanged throughout all evolution steps, serving as the local capacity limit of the cell in the corresponding category. The evolution steps are repeated until the maximum absolute value of the change in the category resource index values ​​of all cells in all mobile categories between two consecutive evolution steps is less than the preset convergence threshold, at which point the evolution stops.

[0011] Furthermore, the leader reads the resource candidate allocation status, and for each cell, takes the minimum difference between the resource index value of each of the five categories and the corresponding task index value as the cell's comprehensive resource sufficiency. All cells are sorted from smallest to largest comprehensive resource sufficiency and assigned a regional priority number starting from 1. Within each cell, each resource node is sorted from largest to smallest resource schedulability and assigned a resource item number starting from 1. The regional priority number and the resource item number are combined into a secondary priority scheduling number and announced to all followers.

[0012] Furthermore, each follower aims to maximize the on-time performance of its assigned flights. For each flight, it selects service resources that meet the following criteria in ascending order of secondary priority scheduling number: aircraft type compatibility, service type matching, and an overlap between the available resource window and the flight's on-time schedule that is not less than the minimum continuous operation time of the corresponding service resource category. The leader, aiming to minimize the total delay time for all flights, checks each service resource to see if there are overlapping requests from two or more followers within the extended occupied period (including preparation and evacuation times). For overlapping resource items, the request with the lower secondary priority scheduling number is retained, while the remaining requests are returned to the corresponding follower to select alternatives from the remaining unoccupied, compliant resources of the same category in ascending order of secondary priority scheduling number. Resources; The leader calculates the distribution of resource items returned from conflict in cells and their categories, adjusts the regional priority numbers of cells in the conflict set forward, and adjusts the resource item numbers within the category of the conflict set forward, triggering the next round of follower responses with the updated secondary priority scheduling numbers; this process is repeated until the number of resource items returned from conflict no longer decreases between two consecutive rounds, at which point the game iteration stops; if the follower still has no available resources of the same category that meet the conditions in the final round, the corresponding flight is transferred to the delayed scheduling queue and a manual intervention request is sent to the airport operations control center; the list of all resource requests that have passed conflict detection is merged into the final resource allocation and scheduling scheme, and the corresponding resource scheduling execution instructions are issued to the boarding gate control terminal, the jet bridge drive controller, the shuttle bus on-board scheduling terminal, the ground staff handheld terminal, and the baggage sorting controller according to the final resource allocation and scheduling scheme.

[0013] A dynamic allocation and scheduling method for airport operation service resources includes the following steps: Step 1: Using airport schedulable service resources as resource nodes and flights to be supported as task nodes, establish connection edges between resource nodes and task nodes according to service matching conditions to generate an airport resource-flight service relationship graph. Perform multi-layer multi-head graph attention aggregation on the airport resource-flight service relationship graph to output the resource embedding vector of each resource node and the task embedding vector of each task node, and decode the resource embedding vector and task embedding vector into non-negative classification scheduling indices respectively; Step 2: Discretize the airport planar area into a cellular grid, inject the classification scheduling indices into each cell, and through synchronous iterative evolution, make the scheduling indices of movable category resources diffuse towards spatial equilibrium among adjacent cells, generating a resource candidate allocation situation; Step 3: With the airport management as the leader and each flight support task agent as the follower, the leader generates and iteratively updates the priority scheduling strategy according to the resource candidate allocation situation, and the followers select service resources according to the priority scheduling strategy. After conflict detection and game iteration, the final resource allocation and scheduling scheme is generated and the resource scheduling execution instruction is issued to the execution terminal.

[0014] By adopting the above technical solutions, the present invention has produced the following beneficial effects: By constructing an airport resource-flight service relationship graph and using a graph attention network for multi-layer multi-head aggregation, the present invention can simultaneously model the complex relationships between five categories of service resources (boarding gates, jet bridges, shuttle buses, ground staff groups, and baggage sorting channels) and all flights to be supported within a unified graph structure framework. It uses an attention mechanism to adaptively identify the degree of differentiated influence of different neighboring nodes on the current node, overcoming the limitations of traditional methods that model each type of resource independently and cannot capture cross-category coupling constraints. This provides a high-quality representation foundation that integrates global topological context information for subsequent scheduling decisions. By decoding embedded vectors into categorized non-negative scheduling indices and injecting them into the cellular space, the scheduling indices of movable resources spontaneously diffuse among adjacent cells towards spatial equilibrium using the synchronous iterative evolution mechanism of cellular automata. Simultaneously, the index values ​​of fixed resources remain constant as local capacity limits. This approach effectively distinguishes between the fundamental differences in spatial allocation mechanisms between fixed and movable resources, avoiding the unfeasible problem of dispersing immovable gates and jet bridges across regions as mobile resources. It ensures that the generated resource candidate allocation situation aligns with the airport's physical operation mechanism. By introducing a Stackelberg game framework, with the airport management as the leader and flight support agents as followers, the leader iteratively updates the priority scheduling strategy based on conflict feedback, and the followers reselect service resources according to the updated strategy. This allows the resource allocation process to achieve a balance between overall airport operational efficiency and the flight support needs of each airline, effectively reducing resource occupation time overlap conflicts and improving the rationality and feasibility of scheduling schemes in multi-entity coordination scenarios. Attached Figure Description

[0015] Figure 1 The schematic diagram of the multi-layer multi-head aggregation architecture of graph attention network provided in this embodiment of the invention; Figure 2 This is a schematic diagram illustrating the principle of embedding vector decoding into category-based scheduling indicators provided in an embodiment of the present invention. Figure 3 This is a schematic diagram of the Mohr's neighborhood diffusion principle of a cellular automaton provided in an embodiment of the present invention. Figure 4 The Stackelberg game iterative convergence curve provided for an embodiment of the present invention. Detailed Implementation

[0016] The airport operation service resource dynamic allocation and scheduling system includes: a relationship graph construction module, which uses available airport service resources as resource nodes and flights to be supported as task nodes, establishes connections between resource nodes and task nodes based on service matching conditions, and generates an airport resource-flight service relationship graph; a graph attention network processing module, which performs multi-layer multi-head graph attention aggregation on the airport resource-flight service relationship graph, outputs the resource embedding vector of each resource node and the task embedding vector of each task node, and decodes the resource embedding vector and task embedding vector into non-negative classification scheduling indices; and a cellular automaton algorithm. The first module discretizes the airport's planar area into a cellular grid, injects categorized scheduling indicators into each cell, and through synchronous iterative evolution, diffuses the scheduling indicators of movable category resources among adjacent cells towards spatial equilibrium, generating a resource candidate allocation situation. The second module, a Stackelberg game scheduling module, uses the airport management as the leader and each flight support task agent as the follower. The leader generates and iteratively updates a priority scheduling strategy based on the resource candidate allocation situation, and the followers select service resources according to the priority scheduling strategy. Through conflict detection and game iteration, the final resource allocation scheduling scheme is generated and resource scheduling execution instructions are issued to the execution terminal.

[0017] In a specific implementation scenario, taking a medium-sized hub airport as an example, the airport is equipped with 32 boarding gates, 18 jet bridges, 45 shuttle buses, 12 ground staff teams, and 8 baggage sorting channels, while there are approximately 120 flights awaiting support. To effectively model the relationship between these service resources and flights, it is first necessary to construct a graph structure that fully reflects the correspondence between resource supply capacity and flight support needs.

[0018] When constructing the airport resource-flight service relationship diagram, each schedulable service resource is established as a resource node, and each flight to be supported is established as a task node. Service resources are divided into five categories according to their functional attributes: Category 1 is boarding gates, Category 2 is jet bridges, Category 3 is shuttle buses, Category 4 is ground staff, and Category 5 is baggage sorting lanes. Boarding gates, jet bridges, and baggage sorting lanes are physically immovable and belong to the fixed category of resources; shuttle buses and ground staff can be redeployed within the airport area and belong to the movable category of resources. This classification method has a direct impact on the subsequent evolution of cellular automata because the spatial distribution of fixed category resources does not change due to scheduling strategies, while movable category resources can be redeployed between different areas.

[0019] For each resource node, four attribute information needs to be collected: resource geographic coordinates, current occupancy status identifier, remaining time until the next release, and category number. The resource geographic coordinates are represented in two dimensions using the airport's local coordinate system; for example, the coordinates of a boarding gate are 1250 meters east and 830 meters north. The current occupancy status identifier is a binary identifier, with a value of 0 indicating availability and a value of 1 indicating current occupancy. The remaining time until the next release is measured in minutes; for example, if a jet bridge is currently serving one flight and is expected to release in 25 minutes, the remaining time is recorded as 25. The category number is an integer from 1 to 5, corresponding one-to-one with the five categories mentioned above. These four attributes are arranged in the order of collection to form a 4-dimensional initial feature vector for the resource node. In actual engineering implementation, geographic coordinates occupy two dimensions, the occupancy status identifier occupies one dimension, the remaining time occupies one dimension, and the category number occupies one dimension; therefore, the initial feature vector of the resource node has five dimensions. In an alternative implementation, auxiliary information such as the resource's rated service capacity and the cumulative number of services provided that day can be added to the initial feature vector of the resource node to provide a richer representation basis.

[0020] For each task node, three attribute information needs to be collected: the geographical coordinates of the parking stand to be used by the flight, the number of seats corresponding to the aircraft type, and the estimated arrival or departure time. The geographical coordinates of the parking stand are also represented using the airport's local coordinate system, serving as spatial anchors for subsequent cellular space injection steps, ensuring the task node has a clear location on the two-dimensional plane. The number of seats corresponding to the aircraft type reflects the passenger scale of the flight; for example, a Boeing 737-800 has 189 seats, and an Airbus A320 has 180 seats. A larger number of seats means more resources are needed for various support functions. The estimated arrival or departure time is represented by a minute offset from 00:00 on the current day; for example, 14:35 is represented as 875. These three attributes are arranged in the order of collection to form a 4-dimensional initial feature vector for the task node, where the geographical coordinates of the parking stand occupy two dimensions, the number of seats occupy one dimension, and the time occupies one dimension.

[0021] When establishing connection edges between resource nodes and task nodes, three conditions need to be met simultaneously. The first condition is aircraft type adaptation, that is, the service specifications of the resource node can match the corresponding flight aircraft type of the task node. For example, if a boarding gate is only applicable to aircraft types C and below, then this boarding gate does not establish a connection edge with flights of aircraft types D and E. The second condition is service type matching, that is, the category of the resource node corresponds to the service link required in the flight guarantee process. For example, a shuttle bus only establishes a connection edge with flights that require remote stand shuttle service, and does not establish a connection edge with flights directly docking at the near stand jet bridge. The third condition is sufficient coverage of the time window. Specifically, the overlapping duration between the resource available time window and the flight guarantee time window is not less than the minimum continuous operation duration of the service resources of the corresponding category. Taking the jet bridge as an example, its minimum continuous operation duration is usually set to 30 minutes, covering the whole process of jet bridge docking, passenger boarding and alighting, and jet bridge evacuation; if the available time window of a certain jet bridge is from 14:00 to 15:20, and the jet bridge guarantee time window of a certain flight is from 14:50 to 15:40, the overlapping interval is from 14:50 to 15:20, a total of 30 minutes, which is exactly equal to the minimum continuous operation duration, so the third condition is met. If the overlapping duration is less than 30 minutes, even if the first two conditions are met, no connection edge is established. This design ensures that each connection edge in the graph represents a service relationship that is practically feasible in the time dimension, avoiding false associations that are nominally reachable but actually unable to complete the full operation.

[0022] In an optional implementation manner, the minimum continuous operation duration can be set separately according to different categories. For example, the minimum continuous operation duration of a boarding gate is set to 45 minutes, that of a jet bridge is set to 30 minutes, that of a shuttle bus is set to 15 minutes, that of a ground crew group is set to 40 minutes, and that of a baggage sorting channel is set to 20 minutes. These values can be adjusted according to the actual operation specifications of the airport.

[0023] Combining all resource nodes, all task nodes, and all undirected edges established according to the above three conditions together forms an airport resource - flight service relationship graph. In the above example scenario, 115 service resources form 115 resource nodes, 120 flights form 120 task nodes, and approximately 1400 undirected edges are established after screening by the three conditions. This graph is essentially a heterogeneous bipartite graph, where resource nodes and task nodes belong to two different node types, and connection edges only exist between nodes of different types.

[0024] After obtaining the airport resource-flight service relationship graph, the next step is to input it into a pre-trained graph attention network to extract deep representations of each node. The reason for using a graph attention network instead of a traditional graph convolutional network is that the contribution of different neighboring nodes to the current node varies significantly in airport scheduling scenarios. For example, for a jet bridge resource node, the impact of an imminent and time-sensitive flight task node is far greater than that of a flight task node with several hours remaining until its arrival time. The graph attention network can adaptively assign different aggregation coefficients to different neighbors through an attention mechanism, thereby more accurately capturing this differentiated service association strength.

[0025] Graph attention networks contain sequentially connected... Layered attention layer, where This indicates the total number of network layers. In actual deployment, A value of 3 achieves a good balance between representational capability and computational cost, because three layers of propagation allow each node to aggregate all neighborhood information within three hops of itself, covering the vast majority of meaningful indirect association paths in the airport resource-flight service relationship graph. Each graph attention layer contains... Each attention head operates independently, among which This represents the number of attention heads, which is set to 4 in practice. The purpose of the multi-head mechanism is to allow the network to learn neighborhood aggregation patterns from multiple different subspaces simultaneously, avoiding the omission of some important association patterns due to the randomness of parameter initialization or gradient direction by a single attention head.

[0026] In each graph attention layer, perform the following aggregation operation on each node in the graph. Let the node to be processed be node . ,node The feature vector held in the current layer is ,in It is 1 3D real vector This indicates the dimension of the feature vectors in the current layer. For the first layer, It equals the dimension of the initial feature vector of the node; for subsequent layers, It is equal to the dimension of the output feature vector of the previous layer.

[0027] for Each attention head in the attention heads, with the first attention head as the first attention head. Let's take one attention point as an example to illustrate, where... The value range is 1 to First, traverse the nodes. Let 'node' be any one of the neighboring nodes directly connected by an undirected edge. ,node The currently held feature vector is . Node eigenvectors With nodes eigenvectors Arrange them in order from beginning to end to form one. A concatenated vector of dimension 1. This concatenated vector is then fed into the 1st dimension. Each attention head corresponds to a single-layer feedforward neural network, which contains one... A linear mapping layer with 3D input and 1D output, and a LeakyReLU activation function. The linear mapping layer will... The 3D concatenated vector is mapped to a scalar. The LeakyReLU activation function performs a nonlinear transformation on this scalar, where the slope of the negative half-axis is set to 0.2 and the slope of the positive half-axis is 1. The scalar output after the transformation is denoted as... , that is, the first Each attention node Relative to node The original attention score. The larger the value, the stronger the node. For nodes The stronger the influence, the better. The reason for using LeakyReLU instead of ReLU is that ReLU's output is always zero in the negative interval, which would cause the attention scores of some neighboring nodes to be completely truncated, thus losing the gradient signal. LeakyReLU, on the other hand, retains a small but non-zero gradient in the negative interval, so that all neighboring nodes can continue to participate in parameter updates during training.

[0028] After traversing all nodes After considering all neighbor nodes, the original attention scores for each neighbor node are obtained. Next, softmax normalization is performed on the original attention scores of all neighbor nodes. Specifically, for each neighbor node... Its normalized attention coefficient The calculation method is as follows: Take the natural index for the molecule, and use the node The sum of the original attention scores of all neighboring nodes after taking the natural exponent is used as the denominator, and the result is obtained by dividing the sum by the original attention scores. After softmax normalization, the sum of the normalized attention coefficients of all neighboring nodes is strictly equal to 1, and each coefficient is non-negative, thus transforming the original score into an aggregated weight in the form of a probability distribution.

[0029] refer to Figure 1 ,like Figure 1 As shown, the left side of the diagram represents the input layer, which contains two types of nodes: resource nodes and task nodes. Resource nodes are... , , This indicates that they represent different categories of service resources within the airport; task nodes are... , , The symbols represent the flights to be supported. Undirected edges are established between resource nodes and task nodes based on aircraft type compatibility, service type matching, and sufficient time window coverage, thus forming an airport resource-flight service relationship graph. Each resource node carries an initial feature vector consisting of resource geographic coordinates, occupancy status identifier, remaining release time, and category number. Each task node carries an initial feature vector consisting of parking position geographic coordinates, number of seats, and estimated time. The initial feature dimensions are... The value is 5. The middle area of ​​the graph contains three layers of graph attention layers arranged sequentially, namely... The value is 3. Each layer contains... An independent attention head, as shown in the figure The value is 4, and each value is marked as an attention head. =1 to =4. In each layer, each attention head independently performs an attention-weighted aggregation operation on the neighbor features of each node in the graph, outputting a single-headed aggregation vector. Then, the single-headed aggregation vectors output by the four attention heads are concatenated to form the output feature vector of the current layer, which is then fed into the next layer as input. The diagram illustrates the change in input and output dimensions for each layer: Layer 1 input dimension... =5, output dimension after concatenation of 4 heads. =20; Level 2 input dimension =20, Output Dimensions =80; Level 3 input dimension =80, Output Dimensions =320. The right side of the graph is the output layer. The output of the third layer is branched into resource embedding vectors and task embedding vectors, both with a dimension of 320. The resource embedding vector carries a deep representation of the resource nodes after three layers of multi-head attention aggregation, incorporating global graph topology information. The task embedding vector carries the deep representation of the task nodes. The reason for setting the number of layers to 3 is that three layers of propagation enable each node to aggregate all neighborhood information within 3 hops of itself, covering most meaningful indirect association paths in the airport resource-flight service relationship graph, achieving a balance between representation capability and computational efficiency. The role of the multi-head mechanism is to learn neighborhood aggregation patterns simultaneously from multiple different subspaces, avoiding the omission of key association information by a single attention head.

[0030] After obtaining the normalized attention coefficients, each neighbor node... The currently held feature vector Multiply by the corresponding normalized attention coefficient Then, perform element-wise summation on the product of all neighboring nodes to obtain the nth... Each attention point is a node The generated single-headed aggregate vector The dimension of this vector is They are the same, both are dimension.

[0031] In all After each attention head independently completes the above aggregation operation, the node get Each of the following single-headed aggregation vectors is: , Until This A single-headed aggregation vector is concatenated in the order of its first and last ends to form a single... A dimensional vector, used as a node The output feature vector of the current layer. This output feature vector is then used as a node. The input feature vector when entering the next graph attention layer. Processed layer by layer in this manner, after... After iterative propagation of the layered graph attention layer, the final output feature vector of each node has incorporated structured contextual information from the multi-hop neighborhood.

[0032] In one alternative implementation, the multi-head outputs of the final graph attention layer are not concatenated, but rather fused using an element-wise arithmetic average. This aims to prevent the dimensionality of the output feature vector from expanding with the number of layers. This approach helps reduce the number of parameters and inference latency in subsequent decoding layers, especially in scenarios with a large number of nodes.

[0033] go through After layer processing, each resource node in the th... The feature vector output by the layer is the resource embedding vector, and each task node in the th layer... The feature vector output by the layer is the task embedding vector. The value is 3. With a value of 4 and an initial feature dimension of 5, if each layer uses a concatenation method to fuse multi-head outputs, then the output dimension of layer 1 is 20, the output dimension of layer 2 is 80, and the output dimension of layer 3 is 320. Both the resource embedding vector and the task embedding vector are 320-dimensional real number vectors.

[0034] Resource embedding vectors and task embedding vectors are abstract representations learned by neural networks in a high-dimensional latent space. The numerical values ​​of each dimension do not have direct physical interpretability and cannot be directly used for subsequent resource surplus / deficit judgment and scheduling priority ranking. Therefore, an additional decoding step is needed to transform the high-dimensional latent representation into a non-negative scalar index with explicit scheduling semantics.

[0035] For decoding resource embedding vectors, each vector is input into a pre-trained fully connected resource decoding layer. The input dimension of this layer is equal to the dimension of the resource embedding vector (320), and the output dimension is 1. The ReLU activation function is used. The ReLU activation function truncates the output value to non-negative, as scheduling metrics semantically represent the schedulability of resources, and negative values ​​have no meaningful connotation. After processing by the resource decoding layer, a non-negative scalar is output as the schedulability of the corresponding resource node. A higher schedulability value indicates higher availability and greater scheduling flexibility for the corresponding resource node in the current time period; a value close to 0 indicates that the resource is nearing full load or is about to enter its occupancy period.

[0036] For decoding the task embedding vectors, each task embedding vector is input into a pre-trained fully connected task decoding layer. The input dimension of the task decoding fully connected layer is also 320, but the output dimension is 5, corresponding to the five service resource categories. The ReLU activation function is used to ensure all output values ​​are non-negative. After processing by the task decoding fully connected layer, five non-negative scalars are output, corresponding to the single-category task demand intensity of categories 1 to 5, respectively. The larger the single-category task demand intensity value, the more urgent the corresponding flight's demand for that category of service resource. The reason for decoding the task embedding vector into five categorical scalars instead of a single composite scalar is that the demand for different categories of resources on the same flight is not consistent. For example, the demand intensity of a large wide-body passenger aircraft for jet bridges and baggage sorting channels may be much higher than its demand for shuttle buses. If the five categories are compressed into a single composite scalar, this differentiated information between categories will be lost, making it impossible for subsequent cellular automata evolution stages to accurately determine the local surplus or deficit of each category of resources.

[0037] In an alternative implementation, the resource decoding fully connected layer and the task decoding fully connected layer can adopt a two-layer fully connected structure with one hidden layer instead of a single-layer fully connected structure. The hidden layer dimension is set to 64 and a ReLU activation function is configured to enhance the nonlinear fitting capability of the decoding mapping.

[0038] The graph attention network, resource decoding fully connected layer, and task decoding fully connected layer are trained together as a unified end-to-end model. Training data comes from historical airport operation records; each training sample contains all resource states, all flight information, and the corresponding actual scheduling results at a given time point. The training objective is to minimize the mean squared error between the predicted resource schedulability and the intensity of single-category task demand and the true indicators implied by the actual scheduling results. After training, the model parameters are fixed, and the online inference phase only requires forward propagation to complete the entire calculation process from the original graph structure to the categorized scheduling indicators within milliseconds.

[0039] After completing the embedding, extraction, and decoding of the graph attention network, we have obtained the resource schedulability of each resource node and the single-category task demand intensity of each task node under the five categories. However, these indicators are still scattered across various discrete nodes and have not yet formed a situation map that can reflect the overall spatial pattern of resource supply and demand at the airport. Therefore, it is necessary to map the above-mentioned categorized scheduling indicators onto a continuous two-dimensional spatial grid. By leveraging the local interaction and iterative evolution mechanism of cellular automata, the scheduling indicators of movable category resources can spontaneously diffuse from surplus areas to scarce areas in space, eventually tending towards a spatial equilibrium state, thereby generating a resource candidate allocation situation.

[0040] Using the aforementioned scenario of a medium-sized hub airport, the area covered by the airport apron and terminal is roughly a rectangle 2400 meters east-west and 1600 meters north-south. This rectangular area is discretized using equally spaced square grids with a side length of 50 meters, resulting in 48 columns east-west and 32 rows north-south, totaling 1536 grid cells, each representing a single cell. The choice of grid side length needs to balance spatial resolution and computational efficiency: too large a side length would result in multiple distant resource and task nodes being grouped into the same cell, obscuring local supply and demand differences; too small a side length would result in many cells containing no nodes, creating meaningless empty cells occupying computational resources. A side length of 50 meters, at this airport scale, roughly corresponds to the spatial span of 1 to 2 parking stands, reasonably grouping geographically close resources and flights with direct service competition into the same cell. In one alternative implementation, the grid side length can be adjusted according to the actual size of the airport and the density of parking spaces. For example, it can be set to 80 meters for large hub airports and 30 meters for feeder airports.

[0041] For each cell, based on the spatial extent it covers in the 2D grid, retrieve all resource nodes and all task nodes whose geographic coordinates fall within the cell's coverage area. Some cells may contain multiple resource nodes and multiple task nodes simultaneously, some cells may contain only resource nodes or only task nodes, and some cells located in the runway, taxiway, or green isolation zone may not contain any nodes.

[0042] Next, each cell is assigned a category-specific scheduling index value. For each of the five categories, the resource schedulability of all resource nodes belonging to the corresponding category within the cell's coverage area is summed, and the resulting value is used as the cell's category resource index value for that category. For example, if a cell contains two boarding gate resource nodes with resource schedulability of 0.85 and 0.62 respectively, then the cell's category resource index value for category 1 is 1.47. If no resource node belonging to a certain category exists within a cell's coverage area, then the corresponding category resource index value is set to 0. Similarly, the single-category task demand intensity of all task nodes within the cell's coverage area is summed, and this sum is used as the cell's category task index value for that category. For example, if a cell contains three flight task nodes with single-category task demand intensities of 0.91, 0.44, and 0.73 respectively in category 3, then the cell's category task index value for category 3 is 2.08. If no task node exists within the cell's coverage area, then the category task index values ​​for all five categories are set to 0. After the above assignment process, each cell obtains 5 sets of category resource indicator values ​​and 5 sets of category task indicator values, for a total of 10 values.

[0043] refer to Figure 2 ,like Figure 2As shown, the upper part of the figure illustrates the decoding path of the resource embedding vector. The resource embedding vector is a 320-dimensional real-valued vector output by the graph attention network. The values ​​of each dimension represent abstract representations learned by the neural network in a high-dimensional latent space, lacking direct physical interpretability and therefore unsuitable for subsequent resource availability assessment and scheduling priority ranking. Thus, a decoding step is needed to transform it into a non-negative scalar indicator with explicit scheduling semantics. Specifically, the resource embedding vector is input into a pre-trained fully connected resource decoding layer with an input dimension of 320 and an output dimension of 1. After linear mapping, it is processed by the ReLU activation function. The ReLU activation function truncates the output value to non-negative because resource schedulability semantically represents the availability of resources, and negative values ​​have no reasonable meaning. The processed output is a single non-negative scalar, representing the resource schedulability. The example value in the figure is 0.85, indicating that the corresponding resource node has high availability in the current time period. The lower part of the figure illustrates the decoding path of the task embedding vector. The task embedding vector is also 320-dimensional, input to a pre-trained fully connected layer for task decoding. This layer has an input dimension of 320 and an output dimension of 5. After processing with the ReLU activation function, it outputs five non-negative scalars, corresponding to the single-category task demand intensity of category 1 (boarding gate), category 2 (airbridge), category 3 (shuttle bus), category 4 (ground staff), and category 5 (baggage sorting lane). The example values ​​in the figure are 0.62, 0.91, 0.44, 0.73, and 0.38, respectively. The reason for decoding the task embedding vector into five categorical scalars instead of a single composite scalar is that the demand for different categories of resources on the same flight is not consistent. For example, the demand intensity of large wide-body passenger aircraft for airbridges and baggage sorting lanes is usually higher than that for shuttle buses. If the five categories are compressed into a single composite scalar, the differentiated information between categories will be lost, making it impossible for the subsequent cellular automata evolution stage to accurately determine the local surplus and deficit status of each category of resources.

[0044] While assigning values ​​to resource indicators, it's also necessary to distinguish the contributions of fixed-category resources and movable-category resources to the cell resource indicators. Category 1 boarding gates, Category 2 jet bridges, and Category 5 baggage sorting lanes are fixed-category resources; their physical locations cannot be changed and they should not participate in subsequent cross-cell spatial diffusion. Category 3 shuttle buses and Category 4 ground staff groups are movable-category resources and can be redeployed between different cells. The category resource indicator values ​​of fixed-category resources will remain unchanged throughout the entire evolution process, serving as the local capacity limit of the cell within that category, for constraint verification in the subsequent Stackelberg game scheduling phase. The significance of this distinction lies in the fact that the service capacity of boarding gates, jet bridges, and baggage sorting lanes is inherently bound to their physical installation locations. They cannot be moved from one apron area to another like shuttle buses. Allowing the indicator values ​​of these fixed resources to diffuse between cells would generate a situational distribution that contradicts physical reality, ultimately rendering the scheduling scheme unenforceable.

[0045] After all cells have been assigned values, the synchronous iterative evolution process is initiated. The core mechanism of the evolution draws on the idea of ​​global state convergence driven by local neighborhood interactions in classical cellular automata. However, unlike traditional discrete-state cellular automata, the state variables of the cells here are continuous real values, and the evolution rules are correspondingly switched from Boolean logic to continuous rules based on numerical comparison and quantitative allocation.

[0046] The execution process of each evolution step is as follows. At the beginning of an evolution step, a global snapshot backup of the current category indicator values ​​of all cells is performed. This involves completely copying the category resource indicator values ​​and category task indicator values ​​of all 1536 cells under the five categories into a separate buffer. Subsequent state transition operations are all read and calculated based on the values ​​in the snapshot buffer, rather than based on the real-time values ​​being updated. The necessity of this synchronous update strategy lies in the fact that if the update operations of each cell depend on values ​​that have been modified by other cells within the same evolution step, the processing order of the cells will affect the final evolution result, causing different situational distributions under the same initial conditions and different traversal orders. This uncertainty is unacceptable for the scheduling system. The global snapshot backup ensures that all cells perform parallel calculations based on the global state at the same moment in the same evolution step, eliminating order dependencies.

[0047] Based on the snapshot state, diffusion operations are performed independently only for movable category resources, namely categories 3 and 4. The specific diffusion process is illustrated using the category 3 shuttle as an example. For each cell in the grid, the snapshot category resource index value and snapshot category task index value of the eight neighboring cells within the Moore neighborhood centered on that cell are read under category 3. The Moore neighborhood refers to the annular neighborhood formed by the eight cells directly adjacent to the central cell in the horizontal, vertical, and diagonal directions. For cells located at the grid boundary, the portion of their Moore neighborhood extending beyond the grid boundary is considered non-existent; subsequent operations are performed only on the actually existing neighboring cells.

[0048] After reading the neighborhood information, each neighboring cell is classified and labeled. Neighboring cells whose category resource index value is strictly greater than the category task index value under category 3 are labeled as surplus cells, indicating that there is an oversupply of shuttle resources in the shuttle resource dimension; neighboring cells whose category resource index value is strictly less than the category task index value are labeled as scarce cells, indicating that there is an undersupply of shuttle resources in the shuttle resource dimension; neighboring cells whose category resource index value is equal to the category task index value are neither labeled as surplus nor scarce, and do not participate in the diffusion operation of this evolution step.

[0049] like Figure 3As shown, the left and right figures illustrate the state changes of the cellular grid in the same region before and after one evolutionary step. The grid consists of 49 cells arranged in 7 rows and 7 columns. Each cell is labeled with two values: the upper value R represents the cell's category resource index value under a certain movable category, and the lower value T represents the cell's category task index value under the same category. When R is strictly greater than T, it indicates that the cell's resource supply exceeds demand under the corresponding category, and it is marked as a surplus cell; when R is strictly less than T, it indicates that the resource supply is insufficient to cover demand, and it is marked as a scarce cell. The cell at the center of the left figure is marked with a bold border, with an R value of 2.7 and a T value of 0.9, which is a typical surplus cell. The dashed box surrounding the central cell indicates the Mole's neighborhood centered on the central cell, i.e., the 8 directly adjacent cells in the horizontal, vertical, and diagonal directions. The right figure shows the diffusion result after one evolutionary step. The central surplus cell divides its category resource index value of 2.7 into 9 equal parts, each approximately 0.3. One part is retained for itself, thus reducing the central cell's evolved R value to 0.3. The remaining 8 parts are distributed to scarce cells in the Moore neighborhood in descending order of their category task index values, increasing the R value of each scarce cell accordingly. The arrows in the diagram indicate the diffusion direction of resource index values ​​from the central surplus cell to the surrounding scarce cells. The division into 9 parts, rather than 8, is because the Moore neighborhood contains 8 adjacent positions plus the central cell itself, totaling 9 positions. The surplus cell needs to retain 1 part for itself to maintain basic local resource reserves, preventing excessive output that could cause it to fall from a surplus state to a scarce state and trigger oscillations. The category resource index value of fixed-category resources remains unchanged throughout all evolution steps; only movable-category resources participate in diffusion. After repeated execution of multiple evolution steps, the category resource index values ​​of all cells in the movable category tend to reach spatial equilibrium.

[0050] For each neighboring cell marked as a surplus cell, the following resource index value allocation operation is performed: the category resource index value of the surplus cell under category 3 is equally divided into 9 parts. The reason for dividing it into 9 parts instead of 8 is that the Moore neighborhood contains 8 neighboring positions plus the center position itself, for a total of 9 positions. The surplus cell needs to reserve 1 part for itself to maintain basic local resource reserves, avoiding excessive output that could cause it to fall from a surplus state to a scarcity state, thus triggering oscillations. One part of the 9 parts is reserved for the surplus cell itself, and the remaining 8 parts are allocated sequentially according to the category task index value of each scarcity cell under category 3, from largest to smallest. Scarcity cells with larger task index values ​​indicate a more urgent need for shuttle service, and therefore receive priority in allocation. Each scarcity cell adds all received shares to its own category resource index value under category 3. If there are no scarcity cells in the Moore neighborhood, i.e., all neighboring cells' resource supply under category 3 is not less than demand, the category resource index value of the surplus cell remains unchanged, and no diffusion is performed.

[0051] In one alternative implementation, surplus cells do not necessarily allocate all remaining 8 portions equally to scarce cells. Instead, they allocate only to the top 3 scarce cells with the highest task metric values, with the remaining portions returned to the surplus cells themselves, thus controlling the diffusion rate and range. This limited diffusion strategy is suitable for scenarios where the airport area is large but the total amount of movable resources is relatively limited, avoiding excessive dilution of resource metric values ​​among distant cells.

[0052] The diffusion operation for Category 4 ground staff is completely independent and parallel to that for Category 3 shuttle buses, with identical diffusion rules; only the category index values ​​are replaced with those for Category 4. The two mobile categories do not interfere with each other.

[0053] The category resource index values ​​of fixed-category resources—namely, Category 1 boarding gates, Category 2 jet bridges, and Category 5 baggage sorting channels—remain unchanged throughout all evolutionary steps. This means that fixed-category resources neither output index values ​​to nor receive index values ​​from their neighbors; their values ​​are locked from the assignment phase. This locking not only conforms to physical reality but also provides a stable capacity boundary reference for subsequent scheduling: when allocating a boarding gate for a flight during the Stackelberg game phase, the fixed-category resource index value of the corresponding cell under Category 1 can be directly read as the local capacity upper limit to determine whether the boarding gate service capacity of that cell region is sufficient to accommodate additional flight support needs.

[0054] Repeat the above evolution steps. As the iterations progress, the index values ​​of movable category resources, initially concentrated in a few cells, gradually diffuse to surrounding scarce areas, and the resource supply and demand difference across the entire grid gradually narrows. The convergence criterion is: between two consecutive evolution steps, the maximum absolute value of the change in category resource index values ​​for all cells across all movable categories is less than a preset convergence threshold, at which point evolution stops. The preset convergence threshold is set to 0.01 in implementation. In the airport scenario described above, convergence is typically achieved after 15 to 25 evolution steps. The set consisting of the 5 sets of category resource index values, the 5 sets of category task index values, and the fixed category local capacity limit for all cells at this point is recorded as the resource candidate allocation status.

[0055] The resource candidate allocation situation is essentially a multi-layered spatial heat map, with each layer corresponding to a service resource category. The color intensity of each cell on the map reflects the balance between the sufficiency of resource supply and the urgency of flight demand in that region within the corresponding category. This situation map provides a spatialized, categorized, and globally balanced decision-making basis for the subsequent game-theoretic scheduling phase.

[0056] After obtaining the candidate resource allocation situation, the process enters the Stackelberg game scheduling phase. The goal of this phase is to transform the spatial supply and demand pattern reflected in the candidate resource allocation situation into a specific, executable resource allocation plan, and to coordinate conflicts among multiple stakeholders.

[0057] Airport operations scheduling involves multiple decision-making entities: airport management, as the manager of infrastructure and public resources, pursues the maximization of overall operational efficiency; while each airline focuses on the quality of service for its own flights. This multi-entity, multi-objective, and information-asymmetric decision-making structure naturally aligns with the modeling framework of Stackelberg games. In Stackelberg games, the leader makes decisions before the followers, and the followers choose their optimal response after observing the leader's decision. The leader anticipates and considers the followers' responses when making decisions. Applying this hierarchical decision-making structure to airport scheduling scenarios allows resource allocation to balance overall efficiency with respect for each airline's specific flight service requirements.

[0058] The airport management is designated as the leader, and the flight support agents of each airline are designated as followers. The leader's utility objective is to minimize the total delay time for all flights, i.e., to minimize the sum of delays caused by waiting for resources. The followers' utility objective is to maximize the on-time performance of their own flights, i.e., each airline agent wants its flights to obtain all the necessary service resources within the planned time as much as possible. There is a correlation between the two levels of utility objectives, but they are not entirely consistent: minimizing the total delay time may require some airlines to make concessions on some flights, and each airline agent, driven by maximizing its own on-time performance, tends to compete for high-quality resources. The resulting competition and conflict are coordinated through game theory iteration.

[0059] like Figure 4As shown, the horizontal axis represents the game iteration rounds, the left vertical axis represents the number of conflict-returned resource items in each category, and the right vertical axis represents the total number of conflict-returned resource items. The figure uses grouped bar charts to show the number of conflict-returned resource items generated in each round of game iteration for category 1 (boarding gates), category 2 (airbridges), category 3 (shuttle buses), category 4 (ground staff), and category 5 (baggage sorting lanes). A line graph shows the total number of conflict-returned resource items across all categories. In round 1, the total number of conflict-returned resource items was 47, with numerous conflicts in each category, reflecting intense resource competition among followers under the initial priority scheduling strategy. Starting from round 2, the leader adjusts the regional priority numbers of cells in the conflict concentration forward and the resource item numbers within the category of the conflict concentration forward, based on the distribution of the cells and categories of the conflict-returned resource items from the previous round. The updated secondary priority scheduling numbers trigger the follower response in the next round. As the game iterates, the total number of conflict-returned resource items decreases round by round, from 47 in round 1 to 28 in round 2, 16 in round 3, 9 in round 4, and 5 in round 5. The number of conflicts of each category also decreases accordingly. In round 6, the total number of conflict-returned resource items is 4, and it remains 4 in round 7. The number of conflict-returned resource items no longer decreases between consecutive rounds, satisfying the termination condition of the game iteration, and the system reaches equilibrium. The remaining 4 conflict-returned resource items in round 7 correspond to flights that cannot find alternative resources of the same category that meet the conditions of aircraft type compatibility, service type matching, and time window coverage among all available resources. These flights will be transferred to the delayed scheduling queue, and a manual intervention request will be sent to the airport operations control center. The overall trend of the curve shows that the mechanism of the leader iteratively updating the priority scheduling strategy based on conflict feedback can effectively guide the resource selection behavior of followers towards coordination, typically resolving major conflicts within 3 to 6 rounds.

[0060] The leader's decision-making process is implemented as follows: The leader reads the resource candidate allocation status and calculates the comprehensive resource sufficiency for each cell. The comprehensive resource sufficiency is defined as follows: In each of the five categories, the difference between the category resource indicator value and the category task indicator value is calculated, and then the minimum value among the five differences is taken as the comprehensive resource sufficiency. The reason for taking the minimum value instead of the average or sum is that flight support is a typical "barrel effect" scenario. The normal support of a flight requires all five categories of service resources to be available. Even if four of the categories are sufficient, if one category is severely deficient, the support of the entire flight will still be hindered. Taking the minimum value can accurately identify the cell region where the resource bottleneck is located, so that the scheduling will prioritize the weakest link.

[0061] Taking a certain cell as an example, the differences in its five categories are as follows: Category 1: 0.32, Category 2: -0.15, Category 3: 0.89, Category 4: 0.41, and Category 5: 0.05. Therefore, the overall resource sufficiency is -0.15. A negative value indicates that the resource supply of this cell is insufficient to cover the demand in at least one category, and this cell belongs to a tense area that needs priority scheduling and attention.

[0062] The leader sorts all cells according to their overall resource abundance, from lowest to highest. A lower overall resource abundance indicates a more severe resource bottleneck in that cell's region, thus placing it higher in the ranking. Following the sorting order, all cells are assigned a region priority number starting from 1 and increasing sequentially. The cell with a region priority number of 1 represents the region with the most acute resource supply-demand imbalance and correspondingly receives the highest scheduling priority.

[0063] After the regional priority number is determined, the leader needs to perform a secondary sorting of resource nodes within each cell. Within each cell, all resource nodes within the cell's coverage area are sorted from highest to lowest resource schedulability, and assigned sequentially resource item numbers starting from 1. Resource nodes with higher schedulability are ranked higher, indicating that the resource is currently in a better available state and is more suitable for priority allocation.

[0064] The regional priority number and resource item number are combined to form a secondary priority scheduling number. When followers select resources, they first select cell regions in ascending order of regional priority number, and then select specific resource nodes in ascending order of resource item number within the same cell. This two-level sorting mechanism ensures that the best resources in the regions with the most prominent resource supply and demand imbalances are allocated first, thereby alleviating bottlenecks at the global level. The leader announces the secondary priority scheduling number to all followers, completing the leader's first round of strategy release.

[0065] After receiving the secondary priority scheduling sequence number, each follower performs resource selection operations for its assigned flights. For each flight under each follower's jurisdiction, the follower selects the required service resources category by category from resources that meet three conditions, in ascending order of the secondary priority scheduling sequence number. The three conditions are consistent with those in the mapping phase: aircraft type compatibility, service type matching, and the overlap between the resource availability time window and the flight support time window is not less than the minimum continuous operation time of the corresponding category of service resources. In addition, for the selected fixed category of resources, it is also necessary to verify whether the local capacity limit of the cell containing the resource under the corresponding category can accommodate the support needs of the corresponding flight. For example, if the local capacity limit of a cell under category 1 corresponds to the service capacity of 2 boarding gates, and 2 flights in this cell have already been assigned boarding gates, then even if the 3rd flight should select a boarding gate in this cell according to the sorting, it must be skipped due to the capacity limit limitation, and the search continues for the next resource that meets the conditions. Each follower summarizes the selection results of all flights to form a follower resource request list and submits it to the leader.

[0066] After the leader compiles a list of all follower resource requests, the conflict detection phase begins. Each service resource is checked item by item to see if there are any overlaps between requests from two or more followers within the extended occupancy period, which includes preparation and evacuation times. The extended occupancy period is the complete occupancy interval formed by adding preparation and evacuation times to both ends of the basic service period for a flight. Taking a jet bridge as an example, the basic service period is 40 minutes for passenger boarding and disembarking, the preparation time is 5 minutes for jet bridge docking, and the evacuation time is 5 minutes for jet bridge withdrawal; the extended occupancy period is 50 minutes. If the extended occupancy periods for the same jet bridge requested by two flights overlap in any way, it is considered a conflict.

[0067] For conflicting resource items, the leader retains the request with the smaller second-level priority scheduling number and returns the remaining requests to their respective followers. The returned followers need to select alternative resources from the remaining unoccupied resources of the same category that meet the conditions, in ascending order of second-level priority scheduling number.

[0068] After completing a round of conflict detection and rollback correction, the leader does not simply repeat the same strategy and wait for convergence. Instead, it updates its priority scheduling strategy based on conflict feedback. This is the core manifestation of the leader anticipating follower responses and adjusting its strategy accordingly in Stackelberg games. Specifically, the leader analyzes the distribution of resource items that rolled back in this round, including their cells and categories. If certain cell regions generated a large number of conflict rollbacks in this round, it indicates that the resource competition pressure in these regions was underestimated in the initial ranking. The leader then advances the priority numbers of these conflict-concentrated cells, giving them higher priority in the next round. Similarly, if certain categories frequently experience conflicts within specific cells, the leader advances the resource item numbers within the corresponding categories. The updated secondary priority scheduling numbers are then redistributed to all followers, triggering the next round of follower responses.

[0069] In one specific implementation, the rule for adjusting the region priority number is as follows: Count the number of conflict rollbacks for each cell in the current round, and move cells with a rollback count greater than 0 forward from the current ranking. The forward movement is proportional to the number of rollbacks; for example, if a cell has 3 rollbacks, it moves forward 3 positions. Resource item numbers are adjusted in a similar manner, moving resource nodes within the conflict-affected resource categories forward within each cell. This conflict-feedback-based strategy update allows the leader to learn the competitive behavior patterns of followers round by round, dynamically shifting scheduling priority towards genuine resource contention hotspots.

[0070] The process of repeatedly updating the leader's strategy and responding to followers forms a multi-round game iteration. The termination condition for the iteration is that the game iteration stops when the number of resource items returned from conflict no longer decreases between two consecutive rounds. The logic behind this termination condition is that when the number of conflicts no longer decreases, it means that the current priority scheduling strategy can no longer eliminate more conflicts through further adjustments, and the system has reached an equilibrium state. In the airport scenario described above, the game iteration typically converges within 3 to 6 rounds.

[0071] Even after the game iterations cease, a small number of unresolved conflicts may still exist because the absolute quantity of certain service resources for certain categories is insufficient to meet the needs of all flights at certain times. In such cases, if a follower still cannot find available resources of the same category that meet the requirements in the final round, the resource request for the corresponding flight will be transferred to the delayed scheduling queue, and a manual intervention request will be sent to the airport operations control center. Flights in the delayed scheduling queue will wait for the next scheduling cycle to participate in resource allocation again, or will be handled specially by a human dispatcher based on the actual situation on site.

[0072] All resource request lists that passed conflict detection were merged to form the final resource allocation and scheduling scheme. This scheme clearly records the correspondence between each flight and each allocated service resource, including the specific resource number, service category, and start and end times of the extended usage period. According to the scheme, corresponding resource scheduling execution instructions are issued to the gate control terminal, the jet bridge drive controller, the shuttle bus onboard dispatch terminal, the ground staff handheld terminal, and the baggage sorting controller. Upon receiving the instructions, the gate control terminal updates the flight allocation information on the flight information display screen. The jet bridge drive controller initiates the jet bridge docking operation at the predetermined time. The shuttle bus onboard dispatch terminal receives the dispatch instruction and displays the navigation route to the target parking position. The ground staff handheld terminal pushes work orders including the work content and arrival time limit. The baggage sorting controller activates the sorting channel and configures baggage tag recognition rules during the corresponding time period.

[0073] In one optional implementation, the final resource allocation and scheduling plan can be submitted to the human-machine interface before being issued for execution, and then reviewed and confirmed by the on-duty dispatcher. The dispatcher can manually adjust individual allocation items in the plan, and the adjusted plan is then uniformly issued to each execution terminal. This human-machine collaborative approach combines the efficiency of algorithm optimization with the flexibility of human experience, and is particularly suitable for unconventional operational scenarios such as extreme weather and large-scale flight delays.

[0074] While specific embodiments of the present invention have been described above, those skilled in the art should understand that these specific embodiments are merely illustrative. Those skilled in the art can omit, substitute, and modify the details of the above methods and systems in various ways without departing from the principles and essence of the present invention. For example, combining the above method steps to perform substantially the same function and achieve substantially the same result according to substantially the same method falls within the scope of the present invention. Therefore, the scope of the present invention is defined only by the appended claims.

Claims

1. An airport operation service resource dynamic allocation and scheduling system, characterized in that, include: The relationship graph construction module uses airport schedulable service resources as resource nodes and flights to be supported as task nodes. It establishes connections between resource nodes and task nodes based on service matching conditions, generating an airport resource-flight service relationship graph. The graph attention network processing module performs multi-layer, multi-head graph attention aggregation on the airport resource-flight service relationship graph, outputting resource embedding vectors for each resource node and task embedding vectors for each task node. It then decodes these vectors into non-negative classification scheduling indices. The cellular automaton evolution module discretizes the airport's planar area into a cellular grid, injects classification scheduling indices into each cell, and through synchronous iterative evolution, diffuses the scheduling indices of movable resource categories towards spatial equilibrium among adjacent cells, generating a resource candidate allocation situation. The Stackelberg game scheduling module uses the airport management as the leader and each flight support task agent as a follower. The leader generates and iteratively updates a priority scheduling strategy based on the resource candidate allocation situation, and the followers select service resources according to the priority scheduling strategy. Through conflict detection and game iteration, a final resource allocation scheduling scheme is generated and resource scheduling execution instructions are issued to the execution terminal.

2. The system according to claim 1, characterized in that, Service resources are divided into 5 categories: Category 1 is boarding gates, Category 2 is jet bridges, Category 3 is shuttle buses, Category 4 is ground staff groups, and Category 5 is baggage sorting channels; Categories 1, 2, and 5 are fixed resources, and Categories 3 and 4 are mobile resources.

3. The system according to claim 2, characterized in that, For each resource node, collect its geographical coordinates, current occupancy status identifier, remaining time until the next release, and category number, and arrange them in the collection order to form the initial feature vector of the resource node; For each task node, the geographical coordinates of the parking position to be used by the flight, the number of seats corresponding to the flight type, and the expected arrival or departure time are collected and arranged in the order of collection to form the initial feature vector of the task node. When a resource node and a task node simultaneously satisfy the requirements of aircraft type adaptation, service type matching, and the overlap between the resource availability time window and the flight guarantee time window is not less than the minimum continuous operation time of the corresponding service resource, an undirected edge is established between the two.

4. The system according to claim 3, characterized in that, The graph attention network consists of L layers of graph attention layers connected in series, with each layer containing K independent attention heads. In each layer, for each attention head, the following operation is performed on each node in the graph: traverse all neighboring nodes directly connected to the current node through undirected edges, concatenate the feature vector of the current node with the feature vector of each neighboring node according to their beginning and end, and then feed them into the single-layer feedforward neural network of the corresponding attention head. After linear mapping and LeakyReLU activation, a scalar is output as the original attention score of the corresponding neighboring node. The original attention scores of all neighboring nodes are normalized using softmax to obtain the normalized attention coefficients of each neighboring node. The feature vector of each neighboring node is multiplied by the corresponding normalized attention coefficient and then summed element by element to obtain a single-head aggregate vector. The K single-head aggregate vectors are concatenated end to end to form the output feature vector of the current node in the current layer and passed to the next layer as input. After processing by L layers, the resource embedding vectors of each resource node and the task embedding vectors of each task node are output.

5. The system according to claim 4, characterized in that, Each resource embedding vector is input into a pre-trained fully connected layer for resource decoding. After ReLU activation, it outputs a non-negative scalar as the resource schedulability of the corresponding resource node. Each task embedding vector is input into a pre-trained fully connected layer for task decoding. After ReLU activation, it outputs five non-negative scalars, which correspond to the single-class task demand intensity of the five categories, respectively.

6. The system according to claim 5, characterized in that, The cellular automaton evolution module divides the plane area of ​​the airport apron and terminal building into equally spaced square grids, with each grid cell serving as a cellular unit. For each cellular unit, it retrieves all resource nodes and all task nodes whose geographic coordinates fall within the cell's coverage area. For each of the five categories, it sums the resource schedulability of all resource nodes within the cell's coverage area that belong to the corresponding category as the cellular's category resource index value, and sums the single-category task demand intensity of all task nodes within the cell's coverage area that belong to the corresponding category as the cellular's category task index value.

7. The system according to claim 6, characterized in that, At the start of each evolution step, a global snapshot backup of the current category index values ​​of all cells is performed; Based on the snapshot state, diffusion operations are performed independently only for movable category resources: For each movable category, the snapshot category resource index value and snapshot category task index value of the 8 neighboring cells within the current cell's Mole neighborhood are read. Neighboring cells whose category resource index value is strictly greater than the category task index value are marked as surplus cells, and those whose category resource index value is strictly less than the category task index value are marked as scarce cells. For each surplus cell, its category resource index value in the corresponding movable category is equally divided into 9 parts, with 1 part reserved for the surplus cell itself. The remaining 8 parts are allocated to the scarce cells in descending order of their category task index values ​​in the corresponding mobile category. If the number of scarce cells in the corresponding mobile category within the Moore neighborhood is 0, the category resource index value of the surplus cells in the corresponding mobile category remains unchanged. The category resource index value of the fixed category resource remains unchanged throughout all evolution steps, serving as the local capacity limit of the cell in the corresponding category. The evolution steps are repeated until the maximum absolute value of the change in the category resource index values ​​of all cells in all mobile categories between two consecutive evolution steps is less than the preset convergence threshold, at which point the evolution stops.

8. The system according to claim 7, characterized in that, The leader reads the resource candidate allocation status. For each cell, the minimum difference between the resource index value of each of the five categories and the corresponding task index value is taken as the cell's comprehensive resource sufficiency. All cells are sorted from smallest to largest comprehensive resource sufficiency and assigned a regional priority number starting from 1. Within each cell, each resource node is sorted from largest to smallest resource schedulability and assigned a resource item number starting from 1. The regional priority number and the resource item number are combined into a secondary priority scheduling number and announced to all followers.

9. The system according to claim 8, characterized in that, Each follower aims to maximize the on-time performance of its assigned flights. For each flight, it selects service resources in ascending order of secondary priority scheduling number, ensuring aircraft type compatibility, service type matching, and that the overlap between the available resource window and the flight's on-time schedule is not less than the minimum continuous operation time of the corresponding service resource category, and submits a resource request list. The leader aims to minimize the total flight delay time across the entire airport, and checks each service resource to see if there are any overlapping requests from two or more followers within the extended occupied time period, including preparation and evacuation times. For overlapping resource items, the request with the smaller secondary priority scheduling number is retained, and the remaining requests are returned to the corresponding followers to reselect alternative resources from the remaining unoccupied resources of the same category that meet the conditions, in ascending order of secondary priority scheduling number. The leader counts the distribution of the cells and categories of the conflict-returned resource items, adjusts the regional priority number of the cells in the conflict set forward, and adjusts the resource item numbers within the category of the conflict set forward, triggering the next round of follower responses with the updated secondary priority scheduling number. This process is repeated until the number of conflict-returned resource items no longer decreases between two consecutive rounds, at which point the game iteration stops. If the follower still has no available resources of the same type that meet the conditions in the final round, the corresponding flight will be transferred to the delayed scheduling queue and a manual intervention request will be sent to the airport operations control center. All resource request lists that pass conflict detection are merged into a final resource allocation and scheduling scheme. According to the final resource allocation and scheduling scheme, corresponding resource scheduling execution instructions are issued to the boarding gate control terminal, the jet bridge drive controller, the shuttle bus on-board scheduling terminal, the ground staff handheld terminal, and the baggage sorting controller.

10. A method for dynamic allocation and scheduling of airport operation service resources, characterized in that, Includes the following steps: Step 1: Using the airport's schedulable service resources as resource nodes and the flights to be supported as task nodes, establish connections between resource nodes and task nodes based on service matching conditions to generate an airport resource-flight service relationship graph. Perform multi-layer multi-head graph attention aggregation on the airport resource-flight service relationship graph to output the resource embedding vector of each resource node and the task embedding vector of each task node. Decode the resource embedding vector and task embedding vector into non-negative classification scheduling indices respectively. Step 2: Discretize the airport's planar area into a cellular grid and inject the classification scheduling indices into each cell. Through synchronous iterative evolution, the scheduling indices of movable category resources diffuse among adjacent cells towards spatial equilibrium, generating a resource candidate allocation situation. Step 3: With the airport management as the leader and the flight support task agents as followers, the leader generates and iteratively updates the priority scheduling strategy based on the resource candidate allocation situation. Followers select service resources according to the priority scheduling strategy. Through conflict detection and game iteration, the final resource allocation scheduling scheme is generated and resource scheduling execution instructions are issued to the execution terminal.

Images