An airport cluster integrated scheduling method based on parallel simulation

Abstract

The application relates to the technical field of air transportation management and intelligent decision-making, and specifically provides an airport cluster integrated scheduling method based on parallel simulation, aiming to solve the problems of poor real-time performance, inability to balance multiple targets and lack of decision-making closed loop in the prior art. The core of the method is as follows: a digital twin model of a physical airport cluster is constructed and continuous time sequence evolution is carried out; preset scheduling events are monitored in real time during the time sequence evolution; once an event is triggered, a target scheduling strategy is generated based on the current state through a parallel simulation and decision-making fusion process; finally, the strategy is mapped back to the digital twin model to replace the current logic to continue the time sequence evolution. Through the parallel system architecture of virtual and real interaction, online, real-time, self-adaptive optimization and decision-making support for the operation of the airport cluster are realized, and the scheduling efficiency, fairness and system resilience are significantly improved.

Description

Technical Field

[0001] This invention relates to the field of air transport management and intelligent decision-making technology, and in particular to an integrated scheduling method for airport clusters based on parallel simulation. Background Technology

[0002] With the continued growth in global air transport demand, the capacity bottleneck of a single airport has evolved into a systemic operational challenge for regional airport clusters. In the context of coordinated airport cluster operations, issues such as airspace traffic management, connecting flights, and resource coordination among multiple operators are becoming increasingly complex, and traditional single-airport scheduling models are struggling to cope.

[0003] Currently, the mainstream technical approach for integrated airport cluster scheduling has significant limitations:

[0004] (1) The real-time dilemma of traditional optimization models: Although methods based on exact algorithms such as integer programming and mixed integer programming can construct rigorous mathematical models, the computational complexity of traditional optimization models increases exponentially with the problem size, making it difficult to provide an effective solution within the minute-level time limit required for decision-making.

[0005] (2) Static trade-off between efficiency and fairness: Existing models usually pre-set fixed weights for efficiency and fairness in the objective function. This static trade-off mechanism lacks flexibility, cannot be intelligently adjusted in a dynamically changing operating environment, and is difficult to meet the actual demands of multiple stakeholders.

[0006] (3) The system architecture is open-loop and lacks real-time decision-making capabilities: Existing high-fidelity simulation tools (such as Simmod and CAST) are essentially offline, open-loop evaluation systems in terms of architecture. They cannot establish a real-time data link with the physical operating environment, resulting in a disconnect between the simulation state and the actual operating state; at the same time, their simulation results cannot be directly and automatically fed back to the physical system for execution as control commands. This architectural defect makes it only suitable for post-event analysis and static scheme evaluation, and it cannot form a closed loop of "perception-decision-execution" during operation, thus failing to support the online adaptive optimization of the system.

[0007] To overcome the aforementioned bottlenecks, parallel simulation technology offers an innovative solution. This technology constructs a virtual mirror system that evolves synchronously with the physical airport cluster and interacts bidirectionally, forming a "virtual-real interaction, parallel evolution" decision-making environment capable of carrying, verifying, and optimizing multi-dimensional scheduling information. Within this environment, the system can perform forward-looking simulations and quantitative evaluations of various scheduling strategies and their associated refined scheduling information (such as flight time slots, resource allocation, and route planning), thereby dynamically generating a comprehensive optimal solution containing complete and executable scheduling instructions. This enables the full-cycle, closed-loop temporal evolution and prediction of airport cluster operations, and facilitates the dynamic optimization and continuous iteration of scheduling strategies and underlying scheduling information, providing accurate and reliable information support for real-time intelligent scheduling in complex environments. Summary of the Invention

[0008] To address the aforementioned technical problems, the technical solution adopted by this invention is as follows:

[0009] This invention provides an integrated airport group scheduling method based on parallel simulation, the method comprising the following steps:

[0010] S100, construct a digital twin model that reflects the operational status of the physical airport cluster.

[0011] S200, based on the digital twin model, performs continuous time-series evolution, and monitors in real time whether a preset scheduling event is triggered during the time-series evolution process; the preset scheduling event includes at least one of capacity overload, connecting flight delay, critical resource node overload, or external system instruction.

[0012] S300, when the preset scheduling event is detected, the parallel simulation and decision fusion process is started based on the current time-series evolution state to generate the target scheduling strategy.

[0013] S400, the target scheduling strategy is mapped to the temporal evolution logic of the digital twin model to replace the currently effective scheduling instruction set, and the continuous temporal evolution continues based on the updated temporal evolution logic.

[0014] The present invention has at least the following beneficial effects:

[0015] 1. It achieves real-time and computable scheduling decisions, fundamentally overcoming the computational bottleneck of traditional optimization models.

[0016] Traditional integer programming and other methods face combinatorial explosion problems when applied to large-scale airport clusters, failing to meet minute-level decision-making requirements. This invention solves this problem through two core innovations: First, it employs an event-triggered mechanism to transform complex global continuous optimization into targeted local instantaneous optimization, significantly reducing the solution space; second, it designs a local state observation window for the deep reinforcement learning agent, reducing the high-dimensional state space to a computable range. Experiments show that this method can respond to events such as capacity overruns within seconds, reducing decision-making time from hours in traditional methods to seconds, meeting the real-time requirements of high-intensity operating environments.

[0017] 2. It breaks through the limitations of static trade-offs among multiple objectives and achieves dynamic intelligent equilibrium between objectives such as efficiency and fairness.

[0018] Traditional methods require manually preset fixed weights, making it difficult to adapt to dynamically changing operational situations. This invention transforms the multi-objective optimization problem into a single-objective learning problem for the agent by designing a composite reward signal that integrates displacement penalties, capacity violation penalties, Jain fairness rewards, migration penalties, and system synergy rewards. The deep reinforcement learning agent, through interaction with the environment, can autonomously learn strategies to balance various objectives in different operational scenarios.

[0019] 3. A closed-loop intelligent management paradigm of perception-temporal evolution-decision-execution was constructed, overcoming the defect of offline simulation being disconnected from physical operation.

[0020] Existing simulation tools can only be used for offline verification. This invention, through a parallel simulation architecture, deeply couples a high-fidelity digital twin model with the real-time data stream of the physical system. After the scheduling strategy undergoes parallel temporal evolution, evaluation, and optimization in the virtual space, the optimal solution can be directly mapped back to the physical system for execution, forming a closed loop of continuous optimization. This mechanism of virtual-physical interaction and parallel evolution transforms simulation from a post-event verification tool into a pre-event decision-making core, greatly improving the feasibility of the scheduling scheme and the overall adaptability and robustness of the system.

[0021] 4. Enhanced system-level collaborative resilience and proactive emergency response capabilities.

[0022] Traditional methods are slow to respond to abnormal events such as capacity saturation at a single airport. This invention, utilizing a parallel simulation mechanism, can synchronously evolve multiple contingency strategies (such as efficiency-first, fairness-first, or load balancing) and perform visual comparative analysis at the application layer. This allows decision-makers to proactively assess the long-term impact of different strategies, shifting from reactive response to proactive planning.

[0023] It should be understood that the description in this section is not intended to identify key or essential features of the embodiments of the present invention, nor is it intended to limit the scope of the invention. Other features of the invention will become readily apparent from the following description. Attached Figure Description

[0024] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0025] Figure 1 The flowchart illustrates an integrated airport cluster scheduling method based on parallel simulation, as provided in this embodiment of the invention. Detailed Implementation

[0026] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0027] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the description of this invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0028] It should be noted that some exemplary embodiments are described as processes or methods depicted as flowcharts. Although the flowcharts describe the steps as sequential processes, many of these steps can be performed in parallel, concurrently, or simultaneously. Furthermore, the order of the steps can be rearranged. A process can be terminated when its operation is complete, but it may also have additional steps not included in the figures. A process can correspond to a method, function, procedure, subroutine, subroutine, etc.

[0029] This invention provides an integrated airport cluster scheduling method and system based on parallel simulation, which addresses the problems of poor real-time performance, difficulty in balancing multiple objectives, and lack of decision-making closed loop in existing airport cluster scheduling technologies.

[0030] This invention is achieved through a four-layer dynamic scheduling system, which includes a physical entity layer, a data layer, a digital twin layer, and an application layer. Its core lies in realizing a continuous self-optimization closed loop in the digital twin layer through reinforcement learning, and finally realizing intelligent decision-making and interaction with the physical world through the application layer.

[0031] Physical entity layer: Serving as the system's sensing unit and execution terminal. It collects real-time operational data (such as location and status) of physical entities like airports, flights, and vehicles through sensors, while simultaneously receiving and executing the final optimized time slot allocation decisions issued by the application layer.

[0032] Data Layer: Serving as the system's data core and interaction hub, it is responsible for collecting, preprocessing, fusing, and storing massive, multi-source, and heterogeneous data from the physical entity layer and external systems (air traffic control, meteorology). At this layer, analysis and prediction algorithms (such as neural network regression and time series models) and pattern recognition algorithms (such as cluster analysis) deeply process the data, predict key parameters (such as delays and taxiing times), and identify operational patterns, providing high-quality, standardized data services and insight support for upper-level intelligent decision-making. Key functions include:

[0033] Data aggregation: Access to real-time data streams such as flight schedules, radar trajectories, resource status, and weather.

[0034] Data preprocessing: The raw data is parsed, cleaned, and converted into standardized state observation vectors required by the intelligent decision-making module.

[0035] Data storage: Utilize relational databases, non-relational databases, and distributed file systems to efficiently store and manage structured and unstructured data.

[0036] The digital twin layer, serving as the system's decision-making and control center, is responsible for building and dynamically updating the digital twin model, and running simulation and intelligent optimization algorithms on top of it. Architecturally, the digital twin layer consists of a model sublayer and an algorithm sublayer.

[0037] The model sublayer is responsible for creating and maintaining the virtual representation of the airport cluster, and is the structured foundation for all data and logic in the twin layer. It includes:

[0038] High-fidelity 3D physical entity modeling: Utilizing professional game engines (such as Unreal Engine) and 3D modeling software, based on airport CAD drawings, BIM data, and high-precision geographic information, highly realistic and dimensionally accurate 3D models of physical entities such as airport runways, taxiways, terminals, aircraft, and vehicles are created. These models not only include geometric appearance but can also be bound with attributes such as materials and physical properties for immersive visualization and high-precision simulation.

[0039] Two-dimensional physical entity modeling: Using geographic information system tools such as WebGIS, a two-dimensional operational status map of the airport cluster is generated. This model presents the layout and operational status (e.g., occupancy, vacancy) of key elements such as runways, taxiways, and aprons in a concise and symbolic manner, and displays key information such as the location, heading, and flight number of aircraft and vehicles in real time. This model is primarily used for daily operational monitoring and as partial state input for reinforcement learning agents.

[0040] Two-dimensional physical entity modeling serves as a "visual guide" for the digital twin system of the flight area. By leveraging the geographic information processing and intuitive visualization capabilities of WebGIS and other technologies, it integrates and processes the complex and diverse key elements in the system and presents them in a concise, clear, efficient, and practical two-dimensional diagram format, achieving a high-fidelity mapping of the airport's daily operation scenarios.

[0041] Two-dimensional physical entity modeling focuses on using concise lines to outline the contours of airport runways and taxiways, clearly presenting their spatial orientation and connectivity. The runway section is labeled with its name, number, and direction information, while the taxiway section highlights key nodes and turning logic. Relevant physical parameters (such as width and length) are uniformly represented through labels or legends.

[0042] The apron is marked with geometric shapes such as rectangles, and each aircraft stand is identified by a number and distinguished by its function (such as jet bridge stands and remote stands). The connection between the apron, taxiways, and runways is clearly defined, facilitating the understanding of the complete process of aircraft entering and leaving the stands.

[0043] In terms of operational status representation, the system uses different icons or colors to distinguish the real-time status of aircraft (taxiing, takeoff, landing, holding, etc.), with accurate updates to location information and arrows indicating the direction of taxiing paths. Key information such as flight number and aircraft type is presented in the form of labels for easy identification and tracking. At the same time, the position and direction of travel of ground support vehicles such as towing vehicles and shuttle buses are also visualized through icons, with their operational status (busy, idle) simultaneously marked, comprehensively reflecting the real-time operational situation on the ground.

[0044] This model supports multi-level zooming and panning operations, allowing for a macro-level grasp of the overall airport layout as well as focusing on local details such as individual parking stands, enabling full-area, multi-granular visualization and analysis. Entity data modeling: A standardized data model is established to describe the static attributes and dynamic states of all entities in the system (such as airfields, aircraft, vehicles, etc.). For example, the aircraft data model includes the aircraft's unique ID, model, dimensions, performance parameters, and real-time updated dynamic attributes such as position coordinates, speed, heading, and operational phase (taxiing, takeoff, landing, etc.). This data model forms the logical foundation driving the entire twin system's operation.

[0045] Entity data modeling is the fundamental support for building a digital twin system for the flight area. In terms of the identification system, the airfield code has uniqueness and exclusivity, and the airfield name serves as the core identifier. Together, they ensure that target objects can be quickly and accurately located in complex data environments and operational scenarios, guaranteeing the accuracy and efficiency of information exchange.

[0046] In terms of location representation, the runway location information is based on high-precision geographic coordinate systems such as WGS84, achieving seamless mapping between the two-dimensional plane and the three-dimensional virtual space. Through a precise positioning mechanism, the virtual model and the actual runway layout maintain a high degree of spatial consistency, providing a reliable basis for subsequent simulation analysis and actual operational decisions.

[0047] In terms of classification, the model divides airfields into subcategories such as runways, taxiways, and connecting taxiways based on the structural characteristics of airport surfaces. Each subcategory follows corresponding design specifications and usage rules according to its functional positioning. Runways focus on parameters related to takeoff and landing performance (such as length), while taxiways focus on parameters related to taxiing rules (such as width and slope), thereby achieving a structured correspondence between functions and attributes at the data level.

[0048] The aircraft data model is the core data carrier of the flight area digital twin system, maintaining a complete and accurate digital profile for each aircraft in the system. By integrating multiple data sources, this model achieves real-time mapping and continuous tracking of the aircraft's entire lifecycle operational status (from pushback, taxiing, takeoff, cruise, landing to parking), providing a unique and reliable data source for upper-level scenario perception, simulation, and intelligent decision-making. The aircraft ID number is the core identity identifier of the aircraft in the twin system, tightly bound to the airline registration number. Through coding rules and database association logic, it enables precise identification and location of individual aircraft within the vast and complex fleet data. From an operational perspective, aircraft types can be categorized as passenger aircraft, cargo aircraft, general aviation aircraft, etc. Based on aircraft model, aircraft can be further classified as Boeing 747, Airbus A380, etc. Information of different types, models, sizes, performance characteristics, and operating procedures is bound together, stored in a structured form, and dynamically updated. Operational attributes are crucial information reflecting the real-time status of an aircraft. Throughout different operational phases, such as taxiing, takeoff, cruise, landing, and parking, real-time dynamic updates and accurate feedback of the status are essential. Real-time aircraft position coordinates can be updated and fed back frequently and with high precision through a data interaction link deeply integrated with the airport's existing navigation and monitoring systems. Heading is precisely guided based on magnetic azimuth, and speed parameters are strictly defined and monitored in real-time according to different operational phases. This provides data support for applications such as surface taxiing trajectory simulation and refined surface traffic management.

[0049] The algorithm sublayer is the core driver of the Siamese layer, embedding the core intelligence of this invention—a time slot allocation model based on deep reinforcement learning. The entire time slot allocation problem is modeled as a Markov decision process. The core elements of the Markov decision process are as follows:

[0050] State space:

[0051] To address the curse of dimensionality, this invention employs an innovative local state observation window. This window is constructed for a single flight requiring a decision. The local observation window is defined as follows: centering on the flight's current reference time slot, it extends k time units forward and backward, forming an observation range containing 2k+1 consecutive time slots. The current reference time slot can be either the originally planned time slot or the latest allocated time slot. The vector data of the state space includes resource status information for each time slot within the observation range. This resource status information includes at least the remaining capacity information for that time slot, distinguishing between arriving and departing capacity.

[0052] In this invention, a time slot refers to a fixed time window (lasting 5 minutes) allocated at a specific airport for the scheduled takeoff or landing of a flight. It is the basic unit of airport capacity management, determining the number of takeoffs and landings the airport can accommodate within a given time period.

[0053] When an agent needs to make a decision about a flight request, it does not observe the entire state of all 288 5-minute time slots of the airport cluster. Instead, its observation space is limited to a time window with a radius of k (e.g., k=6) centered on the current reference time slot of the flight. When considering migrating the flight to be decided to a candidate airport, a local observation window of the same size is generated for each candidate airport, and the center time slot of this local observation window is offset according to the connecting flight time from the original airport to the candidate airport. This design greatly reduces the input dimensionality, allowing the agent to focus on the most relevant spatiotemporal information, thereby achieving efficient learning.

[0054] Action space:

[0055] The action space of an agent is a composite action space, and the action act in this composite action space is defined as a binary tuple act = (A target , Δt), where: A targetThe target airport is represented by the airport to be selected for the flight. The range of target airport values ​​includes the original airport and one or more alternative airports within the airport group. Δt represents the displacement of the time slot allocated to the flight relative to the flight's current reference time slot. The value of Δt is restricted to a preset discrete set, which includes zero and positive and negative integers. For example, Δt∈{-6, ..., -1, 0, 1, ..., 6} corresponds to an adjustment range of -30 minutes to +30 minutes. This design ensures both decision-making flexibility and meets the feasibility requirements of actual operation.

[0056] Reward signal (i.e., reward function):

[0057] The agent's reward signal is a weighted sum of multiple reward components, which include at least: displacement penalty, capacity violation penalty, fairness reward, and migration penalty.

[0058] The displacement penalty term is constructed as a monotonically increasing function of the displacement of the newly allocated time slot relative to the original planned time slot, and the penalty intensity of the displacement penalty term increases non-linearly with the increase of the displacement. Specifically, the displacement penalty term satisfies the following condition:

[0059] R disp =-α×|tt m | p , where R disp The term represents the displacement penalty, where t is the newly allocated time slot. m The original time slot is α, which is a preset coefficient, and p is a nonlinear factor greater than 1 to enhance the sensitivity to large adjustments.

[0060] The capacity violation penalty is constructed as a penalty function positively correlated with the degree to which the resource demand of each airport in the airport cluster exceeds its corresponding capacity in each time slot. The capacity violation penalty is calculated as follows: for all airports 'a' and all capacity types 'u' in the airport cluster, the portion of demand exceeding capacity in each time slot is accumulated, and a penalty is applied. Specifically, the capacity violation penalty meets the following conditions:

[0061] R cap =-∑ a∈AP ∑ u∈U (max(0, D) at u -C at u ));

[0062] Among them, R cap D is a penalty item for capacity violations. at u and C at uThese represent the demand and capacity of airport a in time slot t for capacity type u (arrival, departure, total). This penalty term forces the agent to prioritize resolving capacity conflicts. AP is the set of airports in the airport group, and U is the set of capacity types.

[0063] The fairness reward term is constructed as a reward or penalty function that is positively correlated with the change in one or more indicators reflecting the fairness of the allocation of scheduling resources among different operating entities. Fairness reward term R fair Calculated based on the instantaneous change of Jain's fairness index J, i.e., R fair =J t -J (t-1) Any action that improves system fairness is encouraged, while actions that do not are suppressed, thus guiding agents to seek a dynamic balance between efficiency and fairness. The Jain fairness index J is calculated by comprehensively considering the flight weights of each operating entity, i.e., the airlines, and their normalized average time slot displacement. ; where v i w represents the normalized average time slot displacement of airline i. i Let i be the number of flights of airline i, used to assign weights to the size of the airline. The value of i ranges from 1 to n, where n is the total number of airlines.

[0064] In this embodiment of the invention, the average time slot displacement is defined as the average of the time slot adjustment magnitudes across all flights affected by scheduling decisions by a flight operating entity (such as an airline). For a single flight, the time slot displacement d is the final allocated time slot t for that single flight. assigned Compared with the original planned time slot t planned The absolute value of the difference, i.e., d=|t assigned -t planned The average time slot displacement of a flight operating entity is the arithmetic mean of the time slot displacements of all affected flights under that operating entity.

[0065] Before calculating the fairness index, the average time slot displacement of all flight operation entities is normalized using the max-min normalization method, mapping it to the [0,1] interval. The weights of the flight operation entities can be configured according to actual operational strategies and needs.

[0066] The migration penalty is configured to be activated when a flight airport migration is triggered, and the penalty value of the migration penalty is positively correlated with the estimated cost of the migration. Migration Penalty R airport It is activated when a flight's airport relocation is triggered, and its penalty value consists of the base relocation cost and a variable cost proportional to the transfer time between the original and target airports, i.e., R. airport =-(λ base +λ dist×δ ori-target )×I(AP target ≠AP orig ), where AP target For the target take-off and landing airport, AP orig The target airport is the original departure and arrival airport. If the target departure and arrival airport is equal to the original departure and arrival airport, it means that the aircraft will take off and land at the original airport; otherwise, the opposite will happen. I() represents the indicator function. I() = 1 when the content in parentheses is true, and 0 otherwise. λ base Based on migration cost, λ dist δ is the time cost coefficient. ori-target This refers to the connection time between the original airport and the destination airport. This design guides the agent to prioritize nearby airports with lower costs when selecting an alternate airport.

[0067] In another embodiment of the invention, the reward signal further includes a load balancing reward item, which is configured to be a reward signal positively correlated with the improvement level of load balancing among airports within the airport cluster. Specifically, the load balancing reward item R... balance The following conditions must be met: R balance =-σ({Util a |a∈AP}), where Util a Let σ be the daily load rate of airport a, and let σ() be used to calculate the standard deviation of the load rates of each airport. The core idea of ​​this reward is that a mature airport cluster system should not have a situation where some airports are extremely busy while others are excessively idle. It incentivizes agents to consider not only whether the capacity of alternate airports is sufficient when making flight migrations, but also whether this decision will make the pressure distribution of the entire airport network more balanced, thereby improving the overall resilience of the system.

[0068] In one specific embodiment, the agent's instantaneous reward R in time slot t t The following conditions must be met: R t =w disp ×R t disp +w cap ×R t cap +w fair ×R t fair +w airport ×R t airport +w balance ×R t balance ;w disp w cap w fair w airport and w balanceThese are the weights of the displacement penalty, capacity violation penalty, fairness reward, migration penalty, and load balancing reward, respectively, R. t disp R t cap R t fair R t airport and R t balance These are the displacement penalty, capacity violation penalty, fairness reward, migration penalty, and load balancing reward for the agent in time slot t, respectively.

[0069] In deep reinforcement learning-based scheduling optimization, the reward signal is the core of guiding the agent's learning behavior strategy. This invention overcomes the limitation of fixed weights in traditional reward signals by introducing a dynamic adaptive weight adjustment mechanism based on real-time operating status. This enables the agent to adaptively adjust its optimization focus according to different operating scenarios, thereby making smarter and more rational decisions in complex airport cluster scheduling environments. The core principle of this mechanism is to transform the weight vector of the reward signal from a preset constant into a dynamic function related to the real-time operating status.

[0070] In this embodiment of the invention, the weight of each reward component in the reward signal can be dynamically adjusted according to the real-time operating status. The adjustment criteria include: the current overall load level of the airport cluster; the historical scheduling fairness of each operating entity; and the impact level of abnormal events.

[0071] The current overall load level of the airport cluster can be comprehensively characterized by the weighted average of the ratio of predicted flight counts to total capacity during key future periods, or by the overall system resource utilization rate. When the current overall load level of the airport cluster is high (e.g., during peak hours): the system prioritizes efficiency. In this case, the weights associated with displacement penalties and capacity violation penalties are automatically increased to severely punish decisions that exacerbate delays and overcapacity operations, forcing agents to take more proactive measures to manage traffic flow. When the current overall load level of the airport cluster is low (e.g., during off-peak hours): the system can focus more on long-term benefits and fairness. In this case, the priority of efficiency-related weights is appropriately reduced, while the weights of fairness rewards and load balancing rewards are relatively increased, encouraging agents to optimize the fairness of resource allocation and load balancing among airports.

[0072] The historical scheduling fairness of each operating entity is measured by calculating the Jain fairness index or Theil index between the resources actually obtained (such as slots and gates) and the resources they deserve within a sliding time window (e.g., the past 24 hours). When the historical scheduling fairness of each operating entity is lower than a preset threshold (i.e., historical scheduling has shown significant unfairness): the system will dynamically increase the weight of the fairness reward item to "correct" the agent, guiding it to prioritize improving the fairness of resource allocation in subsequent decisions and compensating the disadvantaged operating entities. When the fairness is at a good level: the weight can be maintained or slightly reduced, and the optimization focus can be shifted to other objectives. Specifically, the "good level" defined in this invention adopts a threshold definition method: when the historical scheduling fairness index (measured by the Jain fairness index) is higher than or equal to a preset threshold, such as 0.9, the system determines that the fairness is at a good level.

[0073] The impact level of abnormal events is determined based on preset abnormal event classification and grading rules, categorizing events such as weather and equipment failures into different impact levels (e.g., Level I / Major, Level II / Severe, Level III / Moderate). When a high-level (e.g., Level I) abnormal event is triggered, the system will enter "Emergency Support Mode." At this time, the weight configuration will be drastically adjusted, with the weight of capacity violation penalties being extremely strengthened (due to a potential sharp drop in capacity), and migration penalties may be temporarily ignored (as large-scale flight relocations become necessary). The core objective is to ensure safety and prevent system crashes. As the event's impact level decreases or dissipates, the weights will smoothly revert to the configuration of the normal operating mode.

[0074] This mechanism, by introducing the aforementioned dynamic weight adjustment strategy, brings about significant technological advancements:

[0075] Context awareness and adaptive optimization: This enables the scheduling system to no longer rigidly execute a single optimization objective, but to perceive the current operational pressure, historical fairness status, and abnormal events, and dynamically adjust its decision preferences, thus achieving true context awareness and adaptive optimization.

[0076] Dynamic trade-offs among multiple objectives: This solves the inherent problem in multi-objective optimization where fixed weights are difficult to adapt to all scenarios. The system can intelligently and dynamically trade off multiple objectives such as efficiency, fairness, cost, and safety, thereby maintaining overall optimality in complex and ever-changing operating environments.

[0077] Enhanced system resilience: When faced with abnormal events, the rapid reconfiguration of weights can guide the system to quickly switch to the most appropriate emergency decision-making mode, greatly enhancing the robustness and resilience of the entire airport cluster dispatch system.

[0078] Through the above-mentioned multi-dimensional and multi-objective reward signal design, the reinforcement learning agent is placed in a complex trade-off environment and guided to learn a comprehensive scheduling strategy that can take into account capacity efficiency, operating costs, airline fairness and overall system stability, so that its decision-making is closer to the complex and ever-changing operational requirements in the real world.

[0079] Network architecture:

[0080] This invention employs Dueling Deep Q-Network (Dueling DQN) as a function approximator to estimate the action value function Q(s, a). The unique feature of the Dueling DQN architecture is that it decomposes the neural network output into two independent branches: one branch estimates the state value function V(s) (representing how good it is to be in a given state s), and the other branch estimates the advantage function A(s, a) (representing how good it is to take an action a in state s relative to the average level). Compared to standard DQN, Dueling DQN can learn the intrinsic value of states more effectively without evaluating the impact of each action. This significantly improves learning efficiency and policy generalization ability in problems with large or redundant action spaces (such as the time slot allocation problem of this invention).

[0081] From a functional perspective, the twin layer mainly includes a simulation engine, a parallel simulation management module, and an intelligent decision-making module. The agent makes decisions in the virtual environment, the simulation engine executes and calculates the reward, and this reward signal is quantified by a series of analytical indicators (such as delay and fairness) and used to update the Q-value function, driving the algorithm to self-optimize and forming an autonomous loop of "decision → simulation → evaluation → optimization".

[0082] Application Layer: Serving as the system's interactive interface and control center, this layer connects internal system functions with external users (such as airport controllers, airline dispatchers, and airport managers) through a collection of interactive interfaces and application services. It provides users with a visual interactive interface and is responsible for task scheduling and final decision-making. Through two-dimensional situation maps and three-dimensional immersive scenes, it intuitively displays the real-time operational status of the airport cluster, the simulation timeline evolution process, and optimization results. Users can perform interactive operations such as zooming, panning, rotating, and querying to gain a comprehensive understanding of the operational situation. It provides operations personnel with intelligent agent-optimized time slot allocation suggestions and highlights potential conflicts, capacity overruns, and other risks. It provides analytical dashboards displaying key performance indicators such as total delays, fairness index, and resource utilization. It allows users to set different hypothetical analysis scenarios in the digital twin environment (e.g., simulating runway closure or thunderstorms in a certain area) and run simulations to evaluate the effectiveness of different emergency plans, providing data support for emergency management and long-term planning.

[0083] Each layer interacts with the others via a communication network and collaborates to complete the following dynamic scheduling method.

[0084] The following is combined with Figure 1 The following details the specific steps for executing the dynamic scheduling method of the present invention through the described system architecture:

[0085] S100, construct a digital twin model that reflects the operational status of the physical airport cluster.

[0086] The simulation engine of the twin layer constructs and maintains a high-fidelity, computable digital twin model based on static configuration data and dynamic real-time data streams obtained from the data layer, accurately reflecting the operational status of the physical airport cluster. The model integrates the following core components:

[0087] Geometric model: Includes the three-dimensional / two-dimensional physical environment of the airport cluster built based on BIM / GIS data.

[0088] Data model: Defines and encapsulates the static attributes and dynamic states of each entity in the system (such as flights and runways).

[0089] Behavioral model: that is, the simulation algorithm, used to simulate the interaction between entities and the system evolution logic.

[0090] The data required to build and drive the digital twin model includes at least the following:

[0091] Static / quasi-static basic data: for example, airport road network topology, fixed set of operating rules (taxiing constraints, runway priority), and route and airspace structure.

[0092] Dynamic real-time data: such as flight schedules and real-time trajectories, dynamic airport capacity values, real-time resource status (occupancy / idleness), and weather information.

[0093] Based on the above inputs, the simulation engine constructs an airport cluster simulation environment that integrates event-driven and discrete-time stepping. This environment can simulate the spatiotemporal evolution of flights throughout the entire process and provide a computational basis for subsequent monitoring and decision-making.

[0094] S200, based on the digital twin model, performs continuous temporal evolution, and monitors in real time whether a preset scheduling event is triggered during the temporal evolution process.

[0095] The simulation engine is based on the digital twin model built by S100 to perform continuous and discrete time step time-series evolution (i.e., deduction) to simulate the entire process of flight operation from pushback, taxiing, takeoff, cruise, landing to parking.

[0096] During this time-series evolution, the monitoring and evaluation module embedded in the simulation engine operates in real time, calculating key performance indicators (such as the slot pressure index SPI) to determine whether a preset scheduling event has been triggered. The preset scheduling events include, but are not limited to, at least one of the following:

[0097] (1) Capacity overrun: The projected flight demand for any airport in any future time slot exceeds the declared or planned capacity of the airport in that time slot;

[0098] (2) Overload of critical resource nodes: The predicted demand of any critical resource node (such as runway, critical taxiway entrance, parking stand area) in any future time slot exceeds the design service capacity threshold of the critical resource node.

[0099] (3) Connecting flight delays: It was detected that the minimum transfer time for subsequent connecting flights may not be met due to the delay of the preceding flight;

[0100] (4) External system instructions: Receive instructions from external systems such as air traffic control and airport operations center. The instructions indicate that a special operational situation has occurred that requires global or local rescheduling (such as sudden weather or equipment failure).

[0101] Furthermore, the preset scheduling event also includes at least one of the following:

[0102] A sudden weather event caused a decrease in airport operational capacity;

[0103] The failure of critical equipment led to a sharp reduction in resource service capabilities;

[0104] Abnormal fluctuations in passenger flow have led to an imbalance in support resources.

[0105] Sudden weather events leading to reduced airport operational capacity include, but are not limited to, severe weather phenomena such as low visibility, strong crosswinds, thunderstorms, snowfall, and freezing rain, resulting in a quantifiable degrade in runway capacity, airspace throughput, or ground operational efficiency at one or more airports. The system integrates meteorological forecasts and real-time observation data. When a specific weather event is predicted or monitored, and its intensity reaches a preset threshold (e.g., visibility below 800 meters or wind speed exceeding 15 meters per second), a scheduling event is automatically triggered. Based on meteorological data and historical operational patterns, the system dynamically estimates and updates the real-time operational capacity of the affected airports, serving as input for subsequent strategy optimization.

[0106] A sharp reduction in resource service capacity caused by critical equipment failure refers to the failure of individual or systemic hardware facilities within an airport cluster that have a decisive impact on operational efficiency. Examples include: runway lighting system malfunction, critical boarding bridge failure, core baggage handling machine shutdown, and primary air traffic control radar failure. The system receives real-time equipment status alarms through interfaces with airport equipment monitoring systems (such as SCADA) or operations management platforms. Once a fault or downtime signal is received from a critical resource node, and the fault is expected to persist for more than a set time threshold (e.g., 10 minutes), a scheduling event is triggered. The system will precisely quantify the degree of reduction in related resource service capabilities (such as runway takeoff and landing rates and gate service speeds) based on the fault type and location.

[0107] Abnormal fluctuations in passenger flow leading to resource imbalances refer to situations where, due to widespread flight delays, diversions, or large-scale events, passenger density in specific airport areas (such as security checkpoints, immigration, and baggage claim areas) far exceeds design capacity within a short period (e.g., a 5-15 minute sliding time window), posing a risk of service unavailability or security risks. Based on multi-source data such as video passenger flow statistics analysis, Wi-Fi / Bluetooth probes, and check-in / security queuing systems, the saturation and queue length of key passenger flow nodes are calculated in real time. When the saturation of any key passenger flow node continuously exceeds a threshold, or a queue length warning is triggered, a scheduling event is initiated. The core optimization objective of this event extends from flight flow to passenger flow, requiring the reallocation of ground handling services, security checkpoints, shuttle buses, and other support resources.

[0108] In this embodiment of the invention, the preset scheduling events are equipped with a priority mechanism to distinguish between local rescheduling and global rescheduling. This mechanism establishes a complete classification and grading system based on the event's spatiotemporal impact, urgency, and impact on the overall system operation, thereby achieving intelligent and differentiated scheduling responses.

[0109] The priority of scheduling events can be divided into high-priority events (triggering global rescheduling) and medium-to-low-priority events (triggering local rescheduling). High-priority events refer to those events that affect multiple airports or have a significant impact on the operational capacity of the core hub of an airport cluster, including but not limited to: large-scale weather deterioration: such as typhoons or widespread thunderstorms covering the core area of ​​the airport cluster; failure of critical shared systems: such as failure of regional navigation facilities or interruption of the primary air traffic control system; and simultaneous capacity crises at multiple airports: a chain reaction causing the predicted demand of multiple major airports to continuously exceed capacity. Such events require resource reallocation and traffic management from the perspective of the entire airport cluster.

[0110] Low-to-medium priority events refer to those whose impact is limited to a single airport or a few flights, and which will not cause a sustained impact on the overall operational stability and capacity of the airport cluster system. Such events include, but are not limited to, the following:

[0111] Short-term capacity overrun at a single airport: Due to short-term, temporary traffic fluctuations, the predicted demand at a single airport may slightly exceed the capacity threshold within a few time slots.

[0112] Temporary overload of critical resource nodes: such as a temporary decrease in the service capacity of a single jet bridge, de-icing pad, or specific taxiway entrance, causing a local resource bottleneck.

[0113] Individual connecting flight connection conflicts: Due to the delay of the preceding flight, one or a few connecting flights may face the risk of insufficient transfer time.

[0114] Limited-terminal equipment failures at a single airport: such as the temporary closure of a runway, part of the baggage conveyor belt, or specific navigation equipment, the impact of which is contained to a local area.

[0115] The core principle for handling such events is local resolution and rapid recovery. By initiating limited optimization algorithms, targeted adjustments are made within the affected airports or resources, aiming to resolve the issue with minimal scheduling intervention and computational overhead. This avoids triggering global, computationally intensive replanning and effectively ensures the overall decision-making efficiency of the system.

[0116] This invention automatically determines priorities and responds through a dedicated event evaluator module. Its specific workflow is as follows:

[0117] Event impact assessment: The system analyzes the spatiotemporal impact of the event in real time to determine whether it involves multiple airports or core hubs;

[0118] Impact Quantification: Quantitatively assess the urgency of the event and the magnitude of the system impact, including indicators such as the degree of capacity loss and the number of affected flights;

[0119] Automatic priority classification: Based on a predefined rule base, the event priority is automatically determined by a comprehensive score that combines the scope of influence and the intensity of the impact.

[0120] Differentiated response triggering: Based on the classification results, intelligently trigger the corresponding range of scheduling responses—global rescheduling or local rescheduling.

[0121] This event evaluator uses a combination of rule-based and machine learning approaches to dynamically optimize judgment accuracy based on historical data and real-time operational status.

[0122] This priority response mechanism achieves precise execution of optimized system resource allocation and scheduling decisions by constructing a precise mapping model between event characteristics and resource impact. Specific technical effects are as follows:

[0123] (1) Optimize the allocation of computing resources and improve the system response efficiency.

[0124] By using an event classifier to perform real-time diagnostics of operational status, the global / local attributes of events can be accurately identified, thereby avoiding the initiation of computationally intensive global optimization algorithms for non-critical events. This mechanism concentrates computing resources on critical events that truly impact system capacity, significantly reducing unnecessary computational load and decision latency.

[0125] (2) Limit the range of disturbance propagation and enhance the stability of the operation plan.

[0126] For localized events, a lightweight optimization algorithm is activated only within the affected airport or resource area to locally absorb and resolve the conflict. This approach minimizes the propagation of disturbances to other established operational plans in the system, maintains the stability of the overall flight schedule and resource allocation scheme, and improves the determinism and predictability of flight execution.

[0127] (3) Achieve precise tiered response and support differentiated decision-making strategies

[0128] By establishing dynamic matching rules between event type, optimization scope, and algorithm strength, differentiated scheduling strategies can be implemented based on the actual impact and severity of events. This mechanism marks the technological evolution of the system from "single global optimization" to "event-driven, hierarchical control," providing core capabilities for airport cluster systems to achieve refined and intelligent scheduling in complex operating environments. This step iteratively executes the following core tasks within each simulation time step:

[0129] Operational simulation: Based on the initial flight schedule set and surface road network structure, drive the simulation model to update the dynamic processes of all aircraft flights, including taxiing, holding, takeoff and landing, and cross-airport transfer flights.

[0130] Resource status update: In each iteration step, the resource status is dynamically updated based on the simulation results, accurately reflecting the current occupied, idle or reserved status of key resources such as runways, taxiways, and parking positions.

[0131] Capacity compliance monitoring: The embedded CapacityMonitor() module is activated, and based on the current and predicted system status, it calculates and monitors the capacity utilization of each airport and each key resource node in real time, and compares it with the preset threshold to provide a basis for judgment for event triggering.

[0132] S300, when the preset scheduling event is detected, the parallel simulation and decision fusion process is started based on the current time-series evolution state to generate the target scheduling strategy.

[0133] When a scheduling event as defined in S200 is detected, the system initiates a parallel simulation and decision fusion process to generate a target scheduling strategy based on the current time-series evolution state. This process consists of the following steps:

[0134] S301: Based on the current time-series evolution state, generate multiple parallel-running computational experimental probes (i.e., simulation instances).

[0135] The parallel simulation management module is activated, and based on the current time-series evolution state, i.e., the complete state of the digital twin model (including flight positions, resource usage, and future predicted states), it synchronously generates multiple parallel computational experiment probes. Each computational experiment probe is an independent copy of the current digital twin model, forming a parallel digital twin environment.

[0136] In this embodiment of the invention, the parallel-running computational experimental probe employs time-compression simulation technology to accelerate temporal evolution in a virtual environment, enabling the evaluation of scheduling strategies for the next several hours within seconds of physical time. The core of time-compression simulation lies in decoupling virtual simulation time from physical system time. By establishing an independent time advancement mechanism within the digital twin model, ultra-real-time temporal evolution of future operating scenarios is achieved. Specific implementation methods include:

[0137] (1) Simulated clock acceleration mechanism:

[0138] In the simulation engine of the digital twin model, set a configurable time compression ratio parameter Rc (Rc is much greater than 1).

[0139] The simulation engine uses the passage of physical system time as a benchmark and advances the virtual simulation time by a factor of Rc. For example, when Rc=100, the virtual simulation time advances by 100 seconds for every 1 second that passes in physical time.

[0140] (2) Model simplification and optimization:

[0141] While ensuring the fidelity of the timing evolution, the simulation model is simplified in stages. High-precision models are used for key core logic (such as flight take-off and landing timing and resource conflict detection), while low-precision models are used or skipped for non-critical auxiliary processes (such as some animation rendering and detailed statistics) to improve speed.

[0142] By utilizing a parallel computing framework, multiple computational experimental probes are distributed across different computing units for simultaneous execution, making full use of the parallel processing capabilities of multi-core processors or computing clusters.

[0143] S302: In each of the computational experimental probes, different optimization criteria are used to perform scheduling optimization and output a set of candidate scheduling strategies.

[0144] By configuring intelligent decision-making modules with different optimization objectives for each computational experimental probe, independent optimization direction exploration is achieved. Specifically, different optimization strategy configurations are achieved by adjusting the weight combination of each component in the reward signal, including but not limited to:

[0145] Efficiency-first strategy: Increase the weight of delay penalty terms, focusing on minimizing the total system delay time;

[0146] Fairness-first strategy: Increase the weight of fairness-based incentives, focusing on improving the fairness of resource allocation among airlines;

[0147] Balanced strategy: Balance the allocation of various weights, and comprehensively consider the operational efficiency and the fairness of resource allocation.

[0148] In at least one computational experimental probe, the scheduling optimization is performed by a trained Deep Q-Network (DQN) agent. This agent takes the current temporal evolution state as input and outputs the optimal scheduling action through its neural network strategy, achieving a preset optimization objective in the corresponding environment, thereby realizing its configured specific optimization criteria within the computational experimental probe.

[0149] After completing the optimization process for a preset time or reaching the convergence condition, each computational experimental probe outputs its corresponding candidate scheduling strategy. The strategies output by all instances together constitute the candidate scheduling strategy set, providing input for subsequent decision fusion.

[0150] Through the above mechanism, a comprehensive exploration from a single optimization objective to multi-objective collaboration has been achieved, ensuring that the final scheduling scheme can adapt to different operating scenarios and optimization needs.

[0151] S303: Through a multi-objective decision fusion mechanism, a target scheduling strategy is determined from a set of multiple candidate scheduling strategies.

[0152] Furthermore, S303 specifically includes the following steps:

[0153] S3031, obtain the comprehensive evaluation value V corresponding to any candidate scheduling strategy, and generate a set of comprehensive evaluation values.

[0154] In this embodiment of the invention, the formula for calculating the comprehensive evaluation value V is: V = k1 × N(E) + k2 × N(R) + k3 × N(F) + k4 × N(C). Where E is the operational efficiency evaluation value, R is the resource utilization evaluation value, F is the fairness evaluation value, C is the cost evaluation value, k1 to k4 are preset weight coefficients that sum to 1, and N(·) is a normalization function used to map the original values ​​of each evaluation value to the [0,1] interval.

[0155] Where, E=α1×(1-T) total / T max )+α2×(1-T avg / Tavg_max )+α3×P punctuality , among which, T total Total system time: The sum of all flight delays (in minutes) within the airport cluster within a preset evaluation time window after executing candidate scheduling strategies in the digital twin model. max Maximum acceptable total delay time threshold: The upper limit of the system's total delay time is set based on historical operational data and operational standards of the airport cluster. avg Average flight delay time: the ratio of the total system delay time to the total number of delayed flights (unit: minutes). T avg_max Maximum average delay threshold: The upper limit for the average flight delay time set based on operational standards. P punctuality Flight punctuality rate: The ratio of the number of flights that actually run on time to the total number of flights within a preset evaluation time window; where on-time operation means that the difference between the actual flight time and the original scheduled time is within a preset tolerance range (e.g., within 15 minutes). Sub-weights for operational efficiency are α1 to α3, and their sum is 1.

[0156] R=β1×U avg +β2×(1–σ Load ), where U avg Runway / parking stand average utilization: the average ratio of actual usage time to total available time for all runways / parking stands within the airport complex during the assessment time window. σ Load Airport cluster load balancing: The standard deviation of the load rate of each airport in the airport cluster within the evaluation time window is used to measure the degree of workload balance among the airports. β1 and β2 are the sub-weights of the resource utilization dimension, and β1 + β2 = 1.

[0157] F=(ΣU j )² / (H×ΣU j ²), where U jThis represents the percentage of resources (e.g., allocated flight slots, parking spaces, and other key resources) that the j-th airline obtains within the airport cluster within a preset evaluation time window. The value of j ranges from 1 to H, where H is the total number of airlines operating flights within the evaluation time window and airport cluster. j This represents the summation of the percentage of resources obtained by H airlines.

[0158] C = 1 - (C reroute +C fuel +C transfer ) / C max , where C reroute Flight rescheduling costs: The total estimated additional costs incurred due to flight rescheduling within the airport cluster as a result of implementing candidate scheduling strategies within a pre-defined assessment time window, including costs for passenger accommodation, ticket rescheduling services, etc. C fuel Additional fuel consumption cost: The total estimated cost of additional fuel consumption (such as in-flight waiting, detours) incurred by flights within the airport group due to the implementation of candidate scheduling strategies within a preset evaluation time window. C transfer Airport relocation cost: The total additional ground support and coordination costs incurred within a pre-defined assessment time window for relocating flights from their original planned airports to alternative airports within the airport cluster due to the implementation of candidate scheduling strategies. C max Maximum acceptable cost threshold: The overall cost ceiling for an airport cluster, set for a single scheduling decision, based on historical data and operational economic requirements.

[0159] For benefit-type indicators (the higher the value, the better), N(X) = (X - Xmin) / (Xmax - Xmin); for cost-type indicators (the lower the value, the better), N(X) = (Xmax - X) / (Xmax - Xmin). In this evaluation model, the operational efficiency evaluation value E, resource utilization evaluation value R, fairness evaluation value F, and cost evaluation value C are all considered benefit-type indicators, meaning that the larger their values, the better the performance in that dimension. The normalization function N(·) uniformly adopts the benefit-type normalization method for all the above indicators.

[0160] The weight coefficients k1 to k4 can be dynamically adjusted according to the operating scenario:

[0161] Typical operating scenario: k1 (efficiency) ≈ 0.4, k2 (resources) ≈ 0.2, k3 (fairness) ≈ 0.2, k4 (cost) ≈ 0.2.

[0162] In peak congestion scenarios: significantly increase the weight of k1 to 0.5-0.6, focusing on efficiency optimization.

[0163] In resource-constrained scenarios: Increase the weight of k2 to 0.3-0.4 to focus on resource balancing.

[0164] Special protection scenario: Increase the weight of k3 to 0.3-0.4 to ensure fairness.

[0165] In this embodiment of the invention, the sub-weight coefficients in the operational efficiency evaluation value E and the resource utilization evaluation value R can be flexibly configured according to the specific optimization orientation of the system. Its design follows the principle of highlighting core contradictions while taking into account auxiliary objectives, aiming to adaptively adjust the optimization focus for different operational scenarios.

[0166] For the efficiency sub-weights (α1, α2, α3):

[0167] α1 (total system delay time weight) is usually assigned the highest priority and configured in the range of 0.4 to 0.7 to force optimization of the overall system smoothness and minimize the total delay of the airport cluster.

[0168] α2 (average delay weight) is configured in a medium range of 0.2 to 0.4 to constrain extreme delays of individual flights and improve operational fairness.

[0169] α3 (on-time performance weight) is set within a basic range of 0.1 to 0.3 as a key performance indicator for the industry to ensure compliance with regulatory and service standards.

[0170] In a preferred embodiment, a configuration of α1=0.5, α2=0.3, and α3=0.2 can be used to achieve a balanced optimization of system efficiency and individual user experience.

[0171] For the resource utilization sub-weights (β1, β2):

[0172] β1 (average utilization weight) is given a higher weight in resource-constrained scenarios, configured in the range of 0.6 to 0.8, in order to maximize the overall utilization efficiency of critical resources (runways, parking spaces).

[0173] β2 (load balancing weight) is increased in scenarios requiring traffic diversion or improved system resilience, and is configured in the range of 0.4 to 0.6 to promote load balancing among units within the airport cluster.

[0174] In a typical configuration, β1=0.6 and β2=0.4 can be used to prioritize resource utilization while giving appropriate consideration to system synergy.

[0175] All sub-weight coefficients satisfy the normalization condition (summing to 1) and are stored in the system as configurable parameters. The system supports dynamic adjustments through the management interface based on actual operating strategies (such as efficiency priority, balance priority, or fairness priority) and specific scenarios (such as peak hours or after abnormal events), thereby giving the scheduling system a high degree of adaptability and strategy flexibility.

[0176] S3032, obtain the candidate scheduling strategy corresponding to the largest comprehensive evaluation value from the comprehensive evaluation value set as the target scheduling strategy (automatic decision-making mode), or present a list of candidate strategies arranged in descending order of comprehensive evaluation value in the visualization interface, supporting decision-makers to finally select the target strategy based on professional experience (human decision-making mode).

[0177] Through the aforementioned multi-objective decision fusion mechanism, this invention achieves the following beneficial effects:

[0178] Scientific decision-making: Through multi-dimensional evaluation, the one-sidedness that may result from optimizing a single objective is avoided, ensuring the comprehensiveness and objectivity of the decision-making basis.

[0179] System adaptability: The dual decision-making path design can not only meet the needs of high-efficiency automation in conventional scenarios, but also meet the needs of human-machine collaboration and experience intervention in complex and special scenarios, which greatly enhances the practicality and flexibility of the system.

[0180] Transparent and Traceable Process: Visualized interactive support and complete decision logs ensure a transparent and trustworthy decision-making process, providing a valuable data foundation for operational review and algorithm iteration. To efficiently handle system-level scheduling events, the system employs an event-type-based adaptive state construction mechanism. This mechanism first identifies key decision-making objects through preprocessing, and then dynamically constructs the most targeted state space for each key decision-making object.

[0181] Furthermore, when any preset scheduling event is detected, a preprocessing procedure is first executed:

[0182] Critical Conflict Identification: Analyze the root causes of events, identify and mark one or more critical flights that cause or may cause resource conflicts or operational risks, and establish the critical flight as a flight to be decided.

[0183] Serial decision initialization: For each flight to be decided, an independent decision process is initialized, and dedicated state space construction, action selection and reward calculation are executed sequentially for that flight.

[0184] For different types of scheduling events, state spaces with different focuses are dynamically constructed for flights to be decided, so as to ensure that the agent obtains the most relevant decision information.

[0185] For capacity overrun events: the state space focuses on the spatiotemporal situation of resource competition. The core state variables of this state space aim to answer where the conflict occurs, with whom it is competing, and how many resources remain, mainly including:

[0186] Capacity saturation: Within a local observation window, the remaining arrival and departure capacity of the target airport and each alternative airport in each time slot.

[0187] Conflicting flight set information: A set of flights that compete for scarce resources and their key attributes (such as aircraft type and operating priority).

[0188] Resource topology status: Real-time and reservation status of affected critical resource nodes (such as runways and taxiway merging points).

[0189] For connecting flight events: the state space focuses on the spatiotemporal continuity of the flight chain. The core state variables of this state space aim to assess whether the preceding flight can be caught, whether the following flight can be waited for, and whether transfer resources are available, mainly including:

[0190] Preceding flight status: its accurate actual / predicted arrival time, current location, and risk of delay.

[0191] Status of subsequent flights: their scheduled departure time and current ground support progress.

[0192] Transit resource status: Real-time availability of key resources such as transit channels and baggage systems.

[0193] Time constraint: The minimum layover time threshold that this connecting flight must meet.

[0194] For external command events: the state space emphasizes global impact and context awareness. The core state variables of this state space aim to understand "what happened, the extent of the impact, and what backup plans are available," mainly including:

[0195] Event context: The type of event described by the instruction (such as weather, fault), the scope of impact, and the expected duration.

[0196] Global impact assessment: Distribution of airports, airspace, and resulting flight backlogs affected by the event.

[0197] Emergency resource status: the availability of special resources such as alternate airports and emergency parking positions.

[0198] Through the above event-driven, on-demand state space design, the system ensures that the reinforcement learning agent can receive the most efficient and relevant information when facing different operational risks, thereby making more targeted and accurate scheduling decisions, and realizing the unification of general decision-making framework and specific scenario optimization.

[0199] In this embodiment of the invention, the training of the deep Q-network agent is an offline learning process based on experience replay and the target network. It learns the optimal time slot allocation strategy through extensive interaction with the simulation engine within the digital twin environment. The specific training process is as follows:

[0200] (1) Interaction and exploration: The agent starts from the initial state and selects actions based on the ε-greedy policy, that is, randomly selects actions with probability ε to explore the environment, otherwise selects the action with the highest value in the current Q value network. This is to balance "exploring" new strategies and "utilizing" existing knowledge.

[0201] (2) Reward calculation and state update: After the environment performs an action, the immediate reward r is calculated based on a comprehensive reward signal. t This function comprehensively considers displacement penalties, capacity violation penalties, fairness rewards, migration penalties, and load balancing rewards. Simultaneously, regarding environmental state updates: if the action involves cross-airport migration, the system will automatically adjust the time slot allocation of related flights to ensure that operational constraints such as minimum turnaround time are met, and update the local observation window to reflect the new state. {t+1} .

[0202] (3) Experience storage: storing the quadruplets (s) generated at each step of the interaction. t a t r t s {t+1} Store it in an experience playback buffer.

[0203] (4) Network update: When the experience pool has sufficient data, randomly sample a batch of data and update the network according to the following steps:

[0204] Calculate the target Q-value: For each sample in the sampling batch, calculate the target Q-value Q based on its recorded immediate reward r and next state s'. target Q target =r+γ×max{a'}Q(s',a';θ), where θ is the parameter of the target network and γ is the discount factor.

[0205] Loss calculation: Calculate the mean squared error loss for all samples in this batch, i.e., the predicted value Q(s, a; θ) of the main network for each sample action and its corresponding target value Q. target The average of the squares of the differences.

[0206] Parameter update: Minimize the above loss using the gradient descent algorithm, backpropagate the error, and update the main network parameters θ.

[0207] Target network synchronization: The main network parameters are slowly synchronized to the target network periodically using a soft update method θ←τθ+(1-τ)θ to improve training stability.

[0208] Policy convergence: Through iterative processes described above, the agent gradually learns a policy that maximizes cumulative rewards. This policy ultimately generates optimized scheduling schemes in real time that simultaneously consider airport capacity constraints, fairness among airlines, and cross-airport collaborative efficiency.

[0209] The Dueling DQN network structure decomposes its output layer into a state value function V(s) and an advantage function A(s, a), and calculates the final Q value through an aggregation layer Q(s, a) = V(s) + (A(s, a) - mean(A(s, a'))). This helps the agent learn the value of a state more efficiently. A(s, a') represents the advantage value corresponding to taking any action a' in state s.

[0210] S400, the target scheduling strategy is mapped to the temporal evolution logic of the digital twin model to replace the currently effective scheduling instruction set, and the continuous temporal evolution continues based on the updated temporal evolution logic.

[0211] This step is crucial for achieving virtual-real interaction and parallel evolution. The specific implementation process is as follows:

[0212] S401, Strategy Issuance and Analysis

[0213] The application layer converts the determined target scheduling strategy into a standardized instruction format and distributes it to the twin layer through the system interface. After receiving the instructions, the parallel simulation management module parses out the specific scheduling operation sequence, including but not limited to: flight time slot reallocation schemes, cross-airport migration instructions, and critical resource rescheduling plans.

[0214] S402, Model State Synchronization

[0215] Before executing policy mapping, the simulation engine first ensures the consistency of the state between the master digital twin model and each parallel simulation environment at the moment the event is triggered. Through a state snapshot mechanism, the key state variables of the current time-series evolution are verified for integrity, providing an accurate baseline state for policy switching. The key state variables of the current time-series evolution include: the precise location and operational status of all flights, the real-time resource occupancy status of each airport, and the airspace traffic distribution.

[0216] In this embodiment of the invention, the integrity verification aims to validate the completeness and consistency of the digital twin model's state snapshot, providing a reliable foundation for seamless policy switching. The verification mainly includes the following four aspects:

[0217] (1) Data completeness verification

[0218] Verify the existence and validity of key status data for all active flights (flights currently in operation or planned to operate), including but not limited to: flight number, precise location (latitude and longitude, specific node on the taxiway or runway), speed, heading, and phase of operation (taxiing, takeoff, landing, holding).

[0219] (2) Verify that the current status (occupied, available, reserved) data of all key resources (runway, taxiway, parking position, boarding gate) has been correctly recorded.

[0220] (3) Logical correlation verification

[0221] Verify that the allocation relationship between flights and resources is correct. For example, verify that a flight marked as "taxiing" has indeed been allocated and associated with a specific taxiway path and its current taxiway segment node.

[0222] Verify whether the planned connection between preceding and following flights remains logically consistent in the model for flight pairs with connecting relationships.

[0223] (4) Timing consistency check

[0224] Verify that all time-related data and events within the model are logically consistent. For example, verify that a flight's estimated departure time should not be earlier than its pushback time, and that the runway time slots it occupies must be within the time period during which it actually uses the runway.

[0225] (5) Resource constraint compliance verification

[0226] Perform a quick static rule check on the current state to ensure there are no obvious, fundamental resource conflicts. For example, verify that the same runway node is not occupied by two or more flights at the same time, or that the same gate is not allocated to multiple flights within the same time period.

[0227] The above verification ensures that the policy switching is based on a real and reliable snapshot of the virtual environment. If the verification fails, an alarm will be triggered and the policy switching process will be paused, rolling back to the previous stable state, or requesting the injection of repaired state data, thereby ensuring the reliability of the entire simulation timing evolution process and the effectiveness of the decision.

[0228] S403, seamless instruction set switching

[0229] The simulation engine injects the new scheduling instruction set into the timing evolution logic in an atomic operation manner, thereby achieving: smooth replacement of the original scheduling rules, maintaining the constraint satisfaction of spatiotemporal continuity, and ensuring the coordinated updating of related flights and resources.

[0230] This process employs a transaction processing mechanism to ensure that if an anomaly occurs during policy switching, it can be rolled back to the previous stable state.

[0231] S404, Continuing with timing evolution and effect verification

[0232] The simulation engine starts from the time breakpoint triggered by the scheduling event and continues to advance the continuous temporal evolution based on the updated temporal evolution logic. During this process:

[0233] Real-time monitoring of the changing trends of various operational indicators under the new strategy;

[0234] Verify whether key constraints such as capacity constraints and inter-trip connectivity are met;

[0235] Compare the degree to which the time-series evolution results match the expected goals.

[0236] S405, Closed-loop feedback establishment

[0237] New data generated during the time-series evolution will be fed back to the system's monitoring module in real time, resulting in: an immediate assessment of the strategy execution effect, the accumulation of experience samples for subsequent optimization iterations, and the improvement of the prediction accuracy of the digital twin model.

[0238] Through the above process, a complete closed loop from strategy generation to execution verification is achieved, ensuring the scientific nature of the scheduling plan and enhancing the system's adaptability through continuous virtual-real interaction. This closed-loop mechanism of decision-making-execution-verification-optimization enables the airport cluster scheduling system to continuously evolve and better cope with complex and ever-changing operating environments.

[0239] After mapping the target scheduling strategy to the temporal evolution logic of the digital twin model, the following steps are also included:

[0240] S500, the target scheduling strategy is converted into standardized control instructions that the physical operating system can directly recognize and execute, and the physical operating system is driven to execute the standardized control instructions to complete the closed loop from virtual simulation decision-making to physical world operation.

[0241] During the instruction standardization process, the optimized scheduling strategy is first parsed into an instruction sequence containing specific operation objects, time parameters, and execution logic. Then, the instruction sequence is converted into a standardized data format (such as using XML or JSON format to encapsulate key parameters) according to the target system interface specification. Finally, the integrity and rationality of the instruction sequence are verified by a logical verification algorithm to ensure that the instruction sequence complies with the safe operation specifications of the physical system.

[0242] After standardization, the system enters the multi-system collaborative distribution phase. Through the integrated multi-protocol communication interface, standardized instructions are distributed in parallel to physical operational systems such as the air traffic control command system, airport operations management system, and airline operations control system. During this process, the system employs a distributed transaction mechanism to ensure synchronous reception and confirmation of instructions across systems, and tracks the instruction distribution status in real time through a status feedback mechanism.

[0243] After the command is issued, the system continuously monitors the execution status. Through the real-time feedback channel established at the data layer, the system monitors the actual execution status of each physical system in response to the command, compares the difference between the expected status and the actual status, and records the timestamps and status change records of the entire process from command issuance and system reception to final execution, forming a complete command traceability chain.

[0244] Finally, the system's execution performance is evaluated and feedback is provided. By comparing key operational indicators (such as flight punctuality rate and resource utilization rate) before and after instruction execution, the degree to which the actual execution performance meets the expected goals is quantitatively assessed. The evaluation results, along with real-time data collected during execution, are fed back to the digital twin layer for calibrating simulation model parameters and optimizing the design of reward signals. This forms a continuous improvement closed loop of decision-making, execution, evaluation, and optimization, continuously improving the accuracy and practicality of subsequent scheduling strategies.

[0245] Through this step, the system achieves a complete closed loop from virtual decision-making to physical execution, ensuring that optimization strategies can be truly implemented and continuously verified and improved in practice. This decision-execution-verification-optimization closed-loop mechanism enables the entire scheduling system to continuously evolve and better adapt to complex and ever-changing operating environments. This physical closed loop, together with the simulation closed loop within the aforementioned digital twin layer, constitutes a complete parallel system architecture, realizing bidirectional interaction and collaborative evolution between the virtual space and the physical world.

[0246] This invention also provides an electronic device, including: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being configured to perform the method described in this invention.

[0247] This invention also provides a computer-readable storage medium storing computer-executable instructions for performing the methods described in this invention.

[0248] It should be understood that the various forms of processes shown above can be used to reorder, add, or delete steps. For example, the steps described in this invention can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution disclosed in this invention can be achieved, and this is not limited herein.

[0249] The specific embodiments described above do not constitute a limitation on the scope of protection of this invention. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this invention should be included within the scope of protection of this invention.

Claims

1. An integrated scheduling method for airport clusters based on parallel simulation, characterized in that, The method includes the following steps: S100, constructing a digital twin model that reflects the operational status of the physical airport cluster; S200, based on the digital twin model, perform continuous temporal evolution, and monitor in real time whether a preset scheduling event is triggered during the temporal evolution process; The preset scheduling events include at least one of the following: capacity overload, connecting flight delay, critical resource node overload, or external system command. S300, when the preset scheduling event is detected, based on the current time-series evolution state, the parallel simulation and decision fusion process is started to generate the target scheduling strategy; S400, the target scheduling strategy is mapped to the temporal evolution logic of the digital twin model to replace the currently effective scheduling instruction set, and the continuous temporal evolution continues based on the updated temporal evolution logic; The parallel simulation and decision fusion process includes: Based on the current time-series evolution state, multiple computational experimental probes that run in parallel are generated simultaneously. Different optimization criteria are used to optimize scheduling in each computational experimental probe, and a set of candidate scheduling strategies is output. A target scheduling strategy is determined from a set of multiple candidate scheduling strategies through a multi-objective decision fusion mechanism. In at least one of the computational experimental probes, the scheduling optimization step is performed by a trained deep Q-network agent; The deep Q-network agent takes the current temporal evolution state as input and outputs scheduling actions to implement the optimization criteria configured for the computational experimental probe in the computational experimental probe. The state space of the deep Q-network agent is a local observation window, which is constructed for a flight to be decided; the local observation window is defined as follows: Centered on the current reference time slot of the flight to be decided, extend k time units forward and backward to form an observation range containing 2k+1 consecutive time slots; The vector data of the state space includes resource status information for each time slot within the observation range. The resource status information includes at least the remaining capacity information for that time slot, and the remaining capacity information distinguishes between inbound capacity and outbound capacity.

2. The method according to claim 1, characterized in that, The action space of the deep Q-network agent is a composite action space, and the action act in this composite action space is defined as the tuple act = (A target ,Δt), where: A target This represents the target airport for the flight to be selected. The range of target airport values ​​includes the original airport and one or more alternative airports within the airport group. Δt represents the displacement of the time slot allocated to the flight to be decided relative to the current reference time slot of the flight. The value of Δt is restricted to a preset discrete set, which includes zero and positive and negative integers.

3. The method according to claim 1, characterized in that, The reward signal of the deep Q network agent is a weighted sum of multiple reward components, which include at least: displacement penalty, capacity violation penalty, fairness reward, and migration penalty.

4. The method according to claim 3, characterized in that, The reward signal also includes a load balancing reward item, which is configured to be a reward signal that is positively correlated with the level of improvement in load balancing among airports within the airport cluster.

5. The method according to claim 1, characterized in that, After mapping the target scheduling strategy to the temporal evolution logic of the digital twin model, the following steps are also included: The target scheduling strategy is converted into standardized control instructions that can be directly recognized and executed by the physical operating system, and the physical operating system is driven to execute the standardized control instructions to complete the closed loop from virtual simulation decision-making to physical world operation.

6. The method according to claim 1, characterized in that, The parallel computing experimental probes employ time compression simulation technology to accelerate the evolution of time sequences in a virtual environment, enabling the evaluation of scheduling strategies for the next several hours to be completed within seconds of physical time.

Images