An interactive collaborative decision-making and autonomous scheduling method for airport ground support vehicles

Abstract

This invention relates to the field of electronic digital data processing, and in particular to an interactive collaborative decision-making and autonomous scheduling method for airport ground support vehicles. The method first acquires and encodes prior knowledge of airport ground support, and then constructs a predictive model for task duration based on historical data. Next, it periodically executes the following steps: acquiring real-time operational visual status including vehicle location, task status, and the set of service objects to be served; using the predictive model to predict the task duration of each service object within the current rolling time window; constructing an optimization model with the objective of minimizing the total service object turnaround delay time, and injecting the predicted task duration as a key parameter into the model; finally, solving the model to generate and output service-oriented resource collaborative orchestration instructions. This invention, by integrating prior knowledge and data intelligence, achieves real-time, automatic, and collaborative scheduling of multiple types of ground support vehicles in dynamic and uncertain environments.

Description

Technical Field

[0001] This invention relates to the field of electronic digital data processing, and in particular to an interactive collaborative decision-making and autonomous scheduling method for airport ground support vehicles. Background Technology

[0002] With the surge in global air travel demand, the ground support systems of large hub airports are facing unprecedented operational pressure. Flight punctuality, a core indicator of airport operational efficiency, heavily relies on the collaborative efficiency of various ground support vehicles (such as refueling trucks, cleaning vehicles, baggage carts, and shuttle buses). However, existing scheduling paradigms generally suffer from the following structural and technical deficiencies when dealing with large-scale, tightly coupled, and highly dynamic complex scenarios:

[0003] 1. The problem of decision silos in resource collaboration

[0004] Most existing systems follow a single vehicle, single flight, or independent scheduling operation model, fragmenting what should be a unified ground support process into multiple independent decision-making silos. This model cannot accurately depict the complex spatiotemporal dependencies and procedural connections between different support services (e.g., refueling is prohibited during passenger boarding and disembarkation, and towing operations can only begin after all preceding flight operations are completed), resulting in suboptimal resource scheduling solutions at the global level and frequent occurrences of vehicle waiting, resource conflicts, and delay propagation.

[0005] 2. Environmental mismatch problem of static scheduling model

[0006] Mainstream research and systems are mostly based on static, deterministic optimization models, whose scheduling schemes are fixed once generated. This sharply contradicts the uncertainties in actual airport operations, such as dynamic adjustments to flight schedules, sudden equipment failures, and rapid changes in weather conditions. Static models lack the ability to respond online and make new decisions; even minor initial disturbances can cause the entire scheduling plan to fail, demonstrating severe environmental mismatch.

[0007] 3. The bottleneck problem of scale effect in algorithm solution

[0008] When the number of flights, vehicles, and missions involved in ground support reaches the daily scale of large hub airports, the decision variable space of the scheduling problem experiences a combinatorial explosion. Exact optimization algorithms (such as mixed integer programming) are computationally too time-consuming in such scenarios and cannot meet the real-time scheduling requirements of minute-level response; while traditional heuristic rules, although fast in solving problems, cannot guarantee the quality of solutions, resulting in a dilemma where efficiency and effectiveness cannot be achieved simultaneously.

[0009] 4. The barriers to integrating knowledge rules and data-driven approaches.

[0010] On the one hand, scheduling logic based on fixed business rules is too rigid and difficult to adapt to changes in operational patterns; on the other hand, purely data-driven predictive models cannot guarantee that scheduling schemes meet all physical and business constraints (such as vehicle capacity and safety intervals). Existing technologies have failed to effectively break down the barriers between prior knowledge (business rules) and data intelligence (operational patterns), resulting in scheduling systems that are either inflexible or lack feasibility.

[0011] Although existing research (such as patent document CN115099615B) has attempted to use multi-agent technologies for collaborative scheduling, the coordination mechanism between agents is still relatively loose, the modeling accuracy for complex temporal and spatial constraints is insufficient, and the computational efficiency and global optimization effect are still challenged when dealing with large-scale real-time scheduling.

[0012] Therefore, there is an urgent need in this field for a new intelligent scheduling paradigm that can fundamentally break down decision silos, achieve deep integration of knowledge and data, and possess online adaptive capabilities, in order to support the efficient and robust operation of large airport ground support systems in dynamic and uncertain environments. Summary of the Invention

[0013] To address the aforementioned technical problems, the technical solution adopted by this invention is as follows:

[0014] This invention provides an interactive collaborative decision-making and autonomous scheduling method for airport ground support vehicles, executed by electronic devices, comprising the following steps:

[0015] S100: Acquire and encode prior knowledge of ground support for the target airport.

[0016] S200, based on the historical mission execution dataset of the target airport, builds a predictive model for mission duration for different types of air traffic operations.

[0017] S300 periodically executes a rolling optimization collaborative orchestration process to generate and output service-oriented resource collaborative orchestration instructions.

[0018] The S300 specifically includes:

[0019] S310, obtain the real-time operational visual status at the current moment. The real-time operational visual status includes the real-time location and availability of all support vehicles, the execution status of all incomplete navigation operations, and the set of service objects that need to be served within the current rolling time window.

[0020] S320, Based on the prediction model, predict the duration of various flight operations for each service object within the current rolling time window, and generate a real-time task duration parameter set.

[0021] S330, based on the prior knowledge and the real-time running visual state, construct an optimization model with the goal of minimizing the total service object turnaround delay time within the current rolling time window; inject the real-time task duration parameter set as a key parameter into the optimization model, and use a metaheuristic optimization algorithm to solve it, generating service-oriented resource collaborative orchestration instructions.

[0022] S340, output the service-oriented resource collaborative orchestration instruction.

[0023] The present invention has at least the following beneficial effects:

[0024] 1. This invention achieves a leap from siloed decision-making to global collaborative control, improving overall system efficiency: By encoding complex prior knowledge such as vehicle capacity, service priority, and spatiotemporal dependencies into constraints of a mixed-integer programming model, a unified service-oriented resource collaborative orchestration model is constructed. This fundamentally solves the resource conflicts and loose coordination mechanisms caused by traditional single-vehicle, single-flight, or independent scheduling modes. This invention can generate globally coordinated service orchestration schemes, significantly reducing flight delays caused by service waiting and resource contention.

[0025] 2. An adaptive immune capability to cope with dynamic uncertainty has been built, ensuring operational robustness: On the one hand, a rolling optimization framework is adopted, enabling service orchestration schemes to be dynamically refreshed and replanned based on real-time perceived operational scenarios (such as flight status and vehicle location), giving the system the agility to respond quickly to emergencies. On the other hand, data-driven modeling is used to predict the duration of service operations at quantile points, transforming uncertainty into robustness parameters and injecting them into the optimization model. This gives the generated schemes inherent buffering capabilities, effectively blocking the chain reaction of delays and improving stability and resilience under interference.

[0026] 3. Overcoming the bottleneck of solution under scale effect and achieving a balance between efficiency and quality: Facing the combinatorial explosion characteristics of the decision space in airport ground resource orchestration problems, this invention adopts an adaptive metaheuristic algorithm (such as an adaptive global optimization engine) as the solution engine within the rolling window. This algorithm efficiently explores the solution space through context-aware intelligent operator selection and a destruction-repair mechanism, finding high-quality, executable orchestration schemes for each rolling window within a limited computation time. It successfully overcomes the shortcomings of exact algorithms in solving large-scale problems and the low solution quality of traditional heuristic algorithms, meeting the stringent requirements of real-time decision-making for large hub airports.

[0027] 4. This invention establishes a closed-loop decision-making process integrating knowledge and data, achieving a significant advancement in scheduling intelligence: It does not simply apply optimization algorithms, but creatively integrates prior knowledge (business rules, physical constraints) with data intelligence (historical patterns, real-time predictions) in a deep and organic manner. Prior knowledge defines the feasible domain boundary for decision optimization, ensuring the feasibility and compliance of the solution; while data intelligence shapes the core of optimization within this boundary, enabling more accurate and intelligent decision-making.

[0028] 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

[0029] 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.

[0030] Figure 1 A flowchart of an interactive collaborative decision-making and autonomous scheduling method for airport ground support vehicles provided in an embodiment of the present invention. Detailed Implementation

[0031] 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.

[0032] 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.

[0033] 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.

[0034] Unless otherwise stated, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art.

[0035] For the purposes of this invention, some key terms are defined as follows:

[0036] Service Object: In an airport ground support scenario, this refers to the entity that needs to receive one or more ground support services. In this invention, the service object specifically refers to flights. As the initiator and destination of all service requests, it is the target entity of the scheduling optimization model. The optimization objective (such as minimizing total turnaround time) directly affects all service objects.

[0037] Service-oriented resources: These refer to standardized service units that, through abstraction and encapsulation, define the specific guarantee capabilities provided by a physical entity as standardized service units that can be uniformly managed, allocated, and invoked. A service-oriented resource is defined by the following elements:

[0038] Service Interface: A clearly defined service type, such as aircraft refueling, cabin cleaning, baggage handling, and passenger shuttle.

[0039] Resource carrier: The physical entity that enables the service capability. For example, a refueling truck is the resource carrier for "aircraft refueling service"; a cleaning team and its equipment are the resource carrier for "cabin cleaning service".

[0040] Status attributes: Information describing the real-time operating status of the resource, including but not limited to its geographical location, availability (idle, busy, unavailable) and current task execution progress.

[0041] For example, refueling truck number 10 at an airport is abstracted as a resource instance in the "refueling service" resource pool. When flight CA101 needs refueling, the scheduling system selects the optimal instance (e.g., truck number 10 after considering location and status) from the "refueling service" resource pool (which may contain multiple refueling trucks) to respond to the service request. Service-oriented resources, as providers of service capabilities, are decision variables in the optimization model for task allocation. The core task of the model is to match the needs of service objects with the optimal service-oriented resource instance.

[0042] Flight operations refer to the smallest logically indivisible unit of action that constitutes a complete ground support service. They form the basis for describing the micro-level execution of a service and constructing precise timing and spatial constraints. They are used for refined modeling in optimization models. By defining dependencies between flight operations (e.g., "connecting the refueling pipeline" must follow "piloting to the aircraft stand"), the feasibility and safety of the generated scheduling scheme are ensured at the micro-level.

[0043] Operational Vision: This refers to the dynamic, digital mapping of the real-time status, future predictions, and interrelationships of all service objects, service resources, and air traffic operations within the entire airport ground support system at a specific decision-making moment. It serves as the unified data foundation upon which the prediction and optimization steps in the rolling optimization cycle rely. It ensures that every decision is based on the latest and most comprehensive system status, and is the core of achieving dynamic response and closed-loop control.

[0044] The Adaptive Global Optimization Engine refers to the core algorithm module used in this invention, which performs real-time solutions within the rolling optimization framework. Its specific implementation is the Adaptive Large Neighborhood Search (ALNS) algorithm. This engine drives the solution to approach the global optimum by iteratively calling deconstruction and reconstruction operation units.

[0045] Operation Unit: In the adaptive global optimization engine, this refers to the basic functional module used for iteratively improving the solution. In this paper, the module that removes some tasks from the current solution is called the deconstruction operation unit; the module that reinserts tasks into the current solution to construct a new solution is called the reconstruction operation unit.

[0046] This invention provides an interactive collaborative decision-making and autonomous scheduling method for airport ground support vehicles. This method integrates prior knowledge and data intelligence, and uses rolling optimization combined with metaheuristic optimization algorithms as the core framework to achieve collaborative scheduling and dynamic optimization of multiple ground support vehicles.

[0047] In specific embodiments, the method provided by the present invention is executed by an electronic device. In one embodiment of the present invention, the electronic device is a computing platform integrating a hardware layer, a system layer, and an application layer, specifically including:

[0048] Hardware entities include at least one central processing unit (CPU) and / or graphics processing unit (GPU) as the core computing power unit; memory for storing various types of data such as operating system, applications, flight data, vehicle status, and rule base; and a communication module for data interaction with airport operation database, vehicle positioning terminal, and vehicle terminal, etc., and its communication methods include but are not limited to wired network, Wi-Fi, 5G / 4G mobile communication and V2X vehicle networking technology.

[0049] Functional modules: The program running on the electronic device is configured as multiple functional modules, including:

[0050] Data acquisition and processing module: used to acquire flight plans, gate information, support vehicle status, location and mission progress information in real time or near real time.

[0051] Optimization Computation Engine: As a core module, it is configured to load and run the rolling optimization framework and metaheuristic optimization algorithm, integrating prior knowledge and real-time data to solve for the optimal or suboptimal vehicle scheduling scheme.

[0052] Dispatch instruction generation and distribution module: used to convert the optimized dispatch scheme into specific, executable instructions, and send them to the corresponding support vehicle or driver terminal through the communication module.

[0053] like Figure 1 As shown in the figure, an interactive collaborative decision-making and autonomous scheduling method for airport ground support vehicles provided by an embodiment of the present invention may include the following steps:

[0054] S100, Prior Knowledge Integration Steps: Acquire and encode prior knowledge of ground support for the target airport.

[0055] In this embodiment of the invention, prior knowledge refers to relatively stable and infrequently changing deterministic information that forms the basis of airport ground support operations. This knowledge defines the boundary and feasible solution space of the scheduling problem and is the core of the "skeleton" for constructing the optimization model. Each time the system re-runs the scheduling algorithm based on real-time data, this framework built from prior knowledge can maintain the stability of the algorithm and ensure that it follows the core operational logic, enabling dynamic optimization to be carried out on a solid foundation.

[0056] Specifically, prior knowledge encompasses multiple dimensions, including but not limited to: ensuring vehicle capacity and speed, airport parking positions and road topology, and the priority and dependency relationships between different air traffic operations.

[0057] S200, Prediction Model Construction Steps: Based on the historical mission execution dataset of the target airport, construct prediction models for mission duration for different types of air traffic operations.

[0058] In this embodiment of the invention, task execution records of the target airport within a set historical period (e.g., 1 month, 3 months, or other time lengths that can reflect the airport's operational patterns) are obtained to form a historical task execution dataset. Each record should include at least: task type (e.g., passenger disembarkation, cabin cleaning, cargo loading, cargo unloading, passenger boarding, refueling, catering, etc.), aircraft type served, planned task start time, actual task start time, and actual task end time. Based on this data, the actual execution time of each task is calculated.

[0059] Optionally, the historical task duration data may be preprocessed before grouping, including but not limited to: handling missing data fields, and identifying and correcting or removing outliers (such as negative durations or extreme values ​​far exceeding the normal range) that deviate significantly from the normal range based on statistical methods (such as box plots based on the 3σ principle) to ensure the quality of the input data.

[0060] Furthermore, the S200 specifically includes:

[0061] S210 groups the historical task execution dataset based on task type and machine model.

[0062] Specifically, the historical task duration data is grouped based on task type and aircraft type as the core dimension, forming multiple homogeneous data subsets such as "A330 aircraft - cabin cleaning" and "B737 aircraft - refueling" to ensure the relevance of subsequent modeling.

[0063] S220: For the task duration data within each group, perform fitting and goodness-of-fit tests on multiple candidate probability distribution models to determine the optimal probability distribution model as the prediction model corresponding to that group.

[0064] Specifically, S220 may include performing the following operations on each data group obtained in S210 to determine its optimal probability distribution model:

[0065] S2201, Candidate Distribution Selection:

[0066] A set of candidate probability distribution models capable of describing positively skewed, continuous random variables is selected for fitting. This invention will consider both parametric and non-parametric distribution methods simultaneously.

[0067] Parameterized distribution: The distribution shape is preset, and the distribution is fitted by estimating parameters. The candidate distributions considered in this invention include:

[0068] Log-normal distribution: suitable for situations where most tasks can be completed on time, but a few tasks have significantly longer durations due to various reasons.

[0069] Gamma distribution: Applicable to the total duration of a task consisting of multiple independent and identically distributed random phases.

[0070] Weibull distribution: Applicable to scenarios where the risk of task completion (i.e., the instantaneous probability of completion) increases or decreases over time.

[0071] Non-parametric distribution: This type of distribution does not pre-determine the distribution shape; estimation is directly driven by the data. In this invention, non-parametric distributions include, but are not limited to, distributions based on kernel density estimation, particularly Gaussian kernel density estimation. Preferably, the bandwidth parameter of the Gaussian kernel density estimation is optimized and determined using the Silverman rule or grid search cross-validation method. This method can adaptively fit the probability density function of the data itself, which may have multi-peak or complex shapes, using a smooth kernel function and optimized bandwidth parameters.

[0072] S2202, Distribution Fitting and Validation:

[0073] This step aims to compare candidate distributions with historical data to select qualified distributions.

[0074] For parameterized distributions: the maximum likelihood estimation method is used to estimate the shape and scale parameters of each distribution and complete the distribution fitting.

[0075] For non-parametric distributions (kernel density estimation): the bandwidth parameter is optimized and determined through methods such as cross-validation to complete the estimation of the probability density function.

[0076] Subsequently, a goodness-of-fit test is performed. This invention primarily employs the Kolmogorov-Smirnov test, which determines whether the distribution hypothesis holds by calculating the maximum difference (D statistic) between the empirical distribution function and the theoretical (or estimated) distribution function of the candidate distribution. A significance level is set (e.g., significance level equal to 0.05). If the p-value of the test is greater than this level, the distribution is considered to have no significant difference from the data, passes the test, and proceeds to the next round of evaluation; otherwise, it is excluded.

[0077] S2203, Determine the optimal distribution:

[0078] Among the qualified distributions that passed the above goodness-of-fit test, in order to further identify the model that best fits the data characteristics and is the simplest, an information criterion is introduced for comprehensive evaluation, including:

[0079] Calculate the Akaike Information Criterion (AIC) and Bayesian Information Criterion (BIC) scores for each qualified distribution. Both AIC and BIC criteria aim to balance the model's goodness of fit with its complexity (usually measured by the number of parameters). The smaller the AIC value, the better the model achieves a balance between accuracy and simplicity.

[0080] By comprehensively comparing the AIC and BIC scores of all qualified distributions, the distribution with the lowest AIC and BIC scores is selected as the optimal probability distribution for this data group, which is the final prediction model. If the optimization results of different criteria conflict, the AIC criterion is given priority because it tends to select the model with stronger predictive power.

[0081] S2304, Quantic Locus Extraction and Application Preparation:

[0082] For each optimal probability distribution model determined in S220, the duration values ​​corresponding to its key quantiles (such as the 90% and 95% quantiles) are extracted. These quantile values ​​will serve as the prediction duration for this type of task at a specific confidence level, directly supporting the real-time prediction step in S320 and providing robust input for subsequent optimization models.

[0083] Furthermore, as a preferred implementation strategy, in the data grouping step of S210, if the number of historical data samples in a group is less than a preset threshold (e.g., 30 samples), the group is considered a small sample group. For small sample groups, they can be merged with other groups of similar task type or machine type to expand the sample size; if merging is not possible, a preset, relatively conservative default duration (e.g., the historical duration limit of this task type across all machine types) is used as its predicted value to ensure the robustness of the scheduling scheme.

[0084] S300 periodically executes a rolling optimization collaborative orchestration process to generate and output service-oriented resource collaborative orchestration instructions.

[0085] Specifically, during airport operating hours, a rolling optimization and collaborative orchestration process is periodically executed based on a preset optimization cycle to generate and output service-oriented resource collaborative orchestration instructions until all flight operations of the service recipients for the day are completed.

[0086] The preset optimization cycle refers to automatically re-executing the entire process from S310 to S340 at a fixed time interval Δt (e.g., 10 minutes or 30 minutes). This cycle Δt is set independently of the length H of the rolling time window and typically satisfies Δt < H. This design ensures that scheduling instructions can be refreshed at a rate higher than that of environmental changes, thereby maintaining the real-time performance and forward-looking nature of the scheduling scheme.

[0087] The termination condition for the process is the completion of all flight operations for the service targets of the day. Specifically, the criteria are: the system detects that the set of service targets to be served within the current rolling time window Fg is empty, and there are no new planned service targets in the airport operations database. This signifies that all ground support operations for the day have been planned, and the rolling optimization scheduling process automatically ends.

[0088] Furthermore, the S300 specifically includes:

[0089] S310, Status Acquisition Step: Acquire the real-time operational visual status at the current moment. The real-time operational visual status includes the real-time location and availability of all support vehicles, the execution status of all incomplete navigation operations, and the set of service objects that need to be served within the current rolling time window.

[0090] In this embodiment of the invention, the current time refers to the absolute system time that triggers the current round of rolling optimization collaborative orchestration. The initial value of the current time is the first time base when the rolling optimization collaborative orchestration process starts each day, and is usually set to a fixed offset (e.g., 2 hours in advance) before the scheduled arrival time of the first scheduled service object at the airport that day. This ensures that the system has generated an initial scheduling plan before the arrival of the first service object.

[0091] In this embodiment of the invention, the real-time running view state is a dynamically updated system snapshot, which is updated based on the scheduling execution results of the previous rolling cycle and the latest external information. The initial state of this real-time running view state is the initial system configuration before any scheduling is performed.

[0092] Specifically, the real-time running visual state includes precisely defined information in the following dimensions:

[0093] (1) Real-time location: refers to the precise geographic coordinates of each support vehicle at the current moment, which is usually obtained through the vehicle-mounted GNSS (Global Navigation Satellite System) module. This location information is mapped onto the airport's digital map to determine the specific functional area where the vehicle is located (such as parking position, driving lane, garage, etc.).

[0094] (2) Real-time availability: refers to the current task status of the vehicle and the expected time when it can be reassigned to a task. Specifically, it is divided into:

[0095] Idle and available: The vehicle has completed its mission and is in standby mode, ready to execute a new mission immediately.

[0096] In progress: The vehicle is currently providing service to a client. At this time, the system records the estimated completion time of its current task, after which the vehicle will become available again.

[0097] Unavailable: The vehicle is temporarily unable to perform its mission due to malfunction, maintenance, or refueling.

[0098] The execution status of all incomplete flight operations:

[0099] This information applies to all assigned but not yet completed flight operations. A status identifier is maintained for each task, specifically including:

[0100] Not Started: The mission has been assigned to a specific vehicle, but the vehicle has not yet arrived at the mission location to begin service.

[0101] In Progress: The vehicle has arrived at the mission location and is performing its service. The system records the actual start time of the mission and dynamically updates its estimated end time based on the predictive model.

[0102] Completed: The task has been completed. This type of task will be removed from the task set for future optimization.

[0103] For tasks that are "not started" or "in progress", they will be inherited into the current rolling time window optimization model, but their time windows, associated vehicles, and other constraints will be updated according to the latest status.

[0104] The set of service objects that need to be served within the current scrolling time window:

[0105] This set defines the scope of service objects that need to be decided in the current optimization cycle, and it is dynamically changing. It includes:

[0106] Planned service objects: All service objects whose planned service time falls within the current rolling time window [t, t+H] according to the service object timetable. t is the current time.

[0107] Added and adjusted service objects: This mainly includes unserved service objects inherited from the previous rolling time window (those postponed due to resource conflicts, etc.), as well as newly added service objects (such as diverted landing service objects) and delayed service objects (whose planned time windows have fallen into the current rolling time window due to previous delays). This dynamic information is obtained in real time from the airport operations database through an interface.

[0108] By integrating the aforementioned multi-dimensional and high-precision real-time status information, accurate and reliable input is provided for subsequent prediction and optimization steps, ensuring that rolling optimization closely follows the dynamic reality of airport operations.

[0109] S320, Real-time prediction step: Based on the prediction model, predict the duration of each flight operation of each service object within the current rolling time window, and generate a real-time task duration parameter set.

[0110] Specifically, the input to this step is the "set of service objects that need to be served within the current scrolling time window" obtained by S310, and the output is the "set of real-time task duration parameters" that can be directly used by the S330 optimization model.

[0111] In this embodiment of the invention, generating the real-time task duration parameter set refers to extracting the quantiles corresponding to a pre-set confidence level from the optimal probability distribution corresponding to the task, and using these quantiles as the predicted duration of the task. This process specifically includes:

[0112] S3201, Task Identification and Model Matching:

[0113] Iterate through each service object within the current rolling time window, identifying each operational task required for that service object (e.g., assigning refueling, cleaning, and catering tasks to service object ABC123). For each task, generate a unique task identifier (e.g., A320-Refueling) based on the task type (e.g., "Refueling") and the aircraft type of the service object (e.g., "A320"). Then, match this identifier with the pre-built optimal probability distribution prediction model for each group in S200 to find the corresponding distribution model and all its parameters (e.g., μ and σ of the log-normal distribution, or bandwidth and kernel function of the kernel density estimation).

[0114] S3202, Quantile Extraction and Duration Prediction:

[0115] For each task, a quantile corresponding to a pre-set confidence level (e.g., 90% or 95%) is extracted from the optimal probability distribution model matched for that task. The value of this quantile is determined as the predicted duration of that flight operation.

[0116] In this embodiment of the invention, selecting a higher confidence level (e.g., 90%) means that there is a 90% probability that the task will be completed within the predicted timeframe during actual execution. This essentially injects a robust buffer into the task plan, allowing more time to cope with random delays in most cases, thereby effectively suppressing the propagation of delays in the scheduling network and improving the stability and reliability of the overall solution.

[0117] S3203, Parameter Set Assembly:

[0118] The predicted durations of all tasks for all service objects within the current rolling time window are aggregated to form a structured real-time task duration parameter set. This parameter set is typically a mapping table or vector, with its core content being {task identifier: predicted duration}. This parameter set will serve as the key input parameter for optimizing the model in step S330.

[0119] Through this step, the present invention transforms the uncertain patterns learned from historical data into specific and actionable deterministic parameters that guide future decision-making, thus achieving a seamless integration of data intelligence and optimized decision-making.

[0120] S330, Model Solving Steps: Based on the prior knowledge and the real-time running visual state, construct an optimization model with the goal of minimizing the total service object turnaround delay time within the current rolling time window; inject the real-time task duration parameter set as key parameters into the optimization model, and use a metaheuristic optimization algorithm to solve it, generating service-oriented resource collaborative orchestration instructions.

[0121] To ensure the smooth operation of ground support activities for service recipients at the target airport, various ground support vehicles must provide a range of services to these recipients within a specified time window. Due to the large number of service recipients and the fact that each recipient requires multiple types of support services, the problem is essentially a multi-vehicle routing problem with resource capacity constraints and time windows. This problem is modeled as an optimization problem defined on an undirected graph G=(V, E). V={0, 1, 2, ..., n} is the set of nodes, where 0 represents a garage node, and F={1, 2, ..., n} is the set of all service recipient nodes; therefore, V={0}∪F. E is the set of edges connecting these service recipient nodes.

[0122] Resources and services are defined as follows:

[0123] There are M different ground support services, each provided by a dedicated heterogeneous fleet. Let Q = {1, 2, ..., M} be the set of these fleets. In one illustrative embodiment, the ground support service may include:

[0124] There is a pre-defined priority relationship between services. For example, when ka < kb (ka, kb∈K), the priority of service ka is lower than that of service kb.

[0125] Each fleet k∈Q contains Nk isomorphic vehicles, which have the same capacity Ck.

[0126] In this embodiment of the invention, the turnaround delay time of the service recipient is the difference between the latest completion time of meal preparation, passenger cabin door closing time, and cargo cabin door closing time and the planned wheel chock removal time.

[0127] In S330, constructing an optimization model based on the prior knowledge and the real-time running visual state includes:

[0128] S3301 sets the initial state for the path constraints of the optimization model based on the real-time location and availability of the vehicle.

[0129] This step aims to transform the system snapshot at rolling moments into the starting point for the optimization model, ensuring that the newly generated scheduling scheme seamlessly integrates with the already completed tasks. Specifically, by injecting the real-time vehicle location and available time as key parameters into the model, which are used to constrain the spatial and temporal starting points of the path, the physical feasibility and temporal consistency of the scheduling instructions are ensured.

[0130] S3302, based on the execution status of unfinished tasks, update the optimization range and boundaries of the current scrolling time window.

[0131] This step automatically includes and removes tasks by dynamically defining the service object node set and node set for the current rolling time window. Simultaneously, the lower bound of the task time window is compared and adjusted with the current rolling time t to ensure that the optimization model only makes decisions for current and future feasible tasks, avoiding ineffective planning for past or completed tasks.

[0132] Furthermore, the constraints followed by the optimization model in S330 include: service uniqueness constraint, route continuity constraint, vehicle usage restriction constraint, garage entry and exit balance constraint, time sequence and continuity constraint, vehicle capacity constraint, service time window constraint, and service priority constraint.

[0133] Based on the updated task set and initial conditions, prior knowledge is concretized into a series of constraints for the model, thereby dynamically instantiating the optimization model of the current window g within the rolling optimization framework.

[0134] Within the scrolling optimization framework, the optimization model for the current scrolling window g is dynamically instantiated. Its node set Vg = {0} ∪ Fg, where Fg is the set of service objects to be served within the current scrolling time window. The specific definitions of the model's objectives and constraints are as follows:

[0135] Objective function: (1);

[0136] Among them, D i This represents the turnaround delay time of service object node i within the current scrolling time window, where i∈Fg.

[0137] Constraints:

[0138] 1. Service uniqueness constraint:

[0139] i∈Fg, k∈Q, i≠j (2);

[0140] Where Vk is the set of vehicles in fleet k, x k ijr Let x be the decision variable. If vehicle r in fleet k travels from service object node i to service object node j, then x k ijr =1, otherwise 0.

[0141] The service uniqueness constraint is used to ensure that each service object is served only once by vehicle r in each fleet k.

[0142] 2. Route continuity constraints:

[0143] , r∈Vk, k∈Q(3)

[0144] , j∈Fg, r∈Vk, k∈Q, u≠j(4)

[0145] Formula (3) ensures that vehicle r cannot have any edge pointing to its current real-time location node z(r). This replaces the assumption in the static model that all vehicles start from the garage, forcing the path to begin from the vehicle's actual location. The decision variable in formula (3) is x. k az（r）r If vehicle r in vehicle k travels from service object node a to service object node z(r), then x k az（r）r =1, otherwise 0. Formula (4) is used to ensure the continuity of the vehicle path, that is, the vehicle must leave after entering the service object node j. The decision variable in formula (4) is x. k ajr and x k jur If vehicle r in vehicle k travels from service object node a to service object node j, then x k ajr =1, otherwise 0. If vehicle r in convoy k travels from service object node j to service object node u, then x k jur =1, otherwise 0.

[0146] 3. Vehicle mission triggering constraints:

[0147] , r∈Vk,k∈Q (5;

[0148] x k z（r）ir Let x be the decision variable. If vehicle r in fleet k travels from service object node z(r) to service object node i, then x k z（r）ir =1, otherwise 0.

[0149] The vehicle task triggering constraint is used to ensure that each vehicle r in each fleet can start at most one new path from its current position node z(r). If a vehicle is still executing a task from the previous window (i.e., is busy), then z(r) is the service object node it is serving, and this constraint will prevent it from being assigned a new task until its status is updated to idle in S310.

[0150] 4. Garage entry and exit balance constraints:

[0151] , r∈Vk, ∀k∈Q(6)

[0152] x k 0ir As a decision variable, if vehicle r in fleet k travels from the garage node to the service object node i, then x k 0ir =1, otherwise 0. k i0r Let x be the decision variable. If vehicle r in fleet k travels from service object node i to garage node, then x k i0r =1, otherwise 0. The garage entry / exit balance constraint ensures that the set of vehicles used in fleet k is balanced within the current rolling window scheduling period. Specifically:

[0153] Left side: Represents the total number of vehicles originating from the starting point. The starting point includes:

[0154] Physical Garage (Node 0): Vehicles currently available in the garage.

[0155] Real-time location z(r): Vehicles that have just completed their mission and are currently located at a certain position within the airport. They will start a new path from their current position.

[0156] Right side: Represents the total number of vehicles that eventually return to the physical garage (node ​​0).

[0157] 5. Time sequence and continuity constraints:

[0158] , i∈Fg,k∈Q (7;

[0159] , j∈Vg, i∈Fg, r∈Vk, k∈Q (8);

[0160] Among them, s k ir s represents the planned start time of service for vehicle r in fleet k to service target node i. k jr t represents the planned start time of service for vehicle r in fleet k to the service target node j. S jk t represents the prediction task duration required for fleet k to provide services to service target node j. T ji x represents the travel time required for a vehicle to travel from service node j to service node i. k jir As a decision variable, if vehicle r in fleet k travels from service object node j to service object node i, then x k jir=1, otherwise, x k jir =0. M is a sufficiently large positive number to ensure that when x = 0. k jir When = 0, the time sequence and continuity constraints are invalid.

[0161] Formula (7) is used to ensure that vehicle r starts service at node i at time s. k ir The time available must be later than the time when the vehicle can start a new task after the current rolling moment, where available(r) is obtained from the system status obtained by S300 and represents the moment when the vehicle finishes the previous task and can be reassigned. If the vehicle is idle, available(r) is equal to the current moment; if the vehicle is in a task, it is equal to its expected task completion time. Formula (8) is used to ensure that if vehicle r travels from node j to node i, the service start time at node i must be later than the service completion time at node j plus the travel time.

[0162] 6. Vehicle capacity constraints:

[0163] , r∈Vk, k∈Q(9)

[0164] Where, d i Let Ck represent the resource requirements of the service target node i, and Ck represent the maximum resource capacity of a single vehicle in fleet k. The vehicle capacity constraint is used to ensure that the amount of resources carried by each vehicle r does not exceed its maximum capacity Ck.

[0165] Vehicle capacity refers to the maximum number of physical resources a support vehicle belonging to a fleet can carry or handle during a single mission. For example, for passenger shuttle buses / crew shuttles, capacity refers to the vehicle's maximum number of passenger seats; the corresponding resource requirement for the service recipient is the number of passengers or crew members the service recipient needs to ferry. For baggage / cargo trailers, capacity refers to the trailer's maximum load capacity or the number of available container / pallet units; the corresponding resource requirement for the service recipient is the total weight of the baggage / cargo or the required number of containers / pallets. For catering trucks, capacity refers to the number of airline meals the truck can carry or the number of catering carts; the corresponding resource requirement for the service recipient is the total number of meals needed by all passengers and crew members on the flight. For water trucks / sewage trucks, capacity refers to the water tank volume (e.g., liters); the corresponding resource requirement for the service recipient is the amount of fresh water added or sewage removed from the service recipient.

[0166] 7. Service time window constraints:

[0167] , i∈Fg, r∈Vk, k∈Q(10)

[0168] , i∈Fg, r∈Vk, k∈Q (11);

[0169] Where t is the current scrolling time, ET i LT is the lower bound of the service time window for the service object node i, i.e., the earliest start time of service. i The upper bound of the service time window for service object node i is the latest start time. The service time window constraint is used to ensure that the service start time of each service object is not earlier than the lower bound of its time window and to ensure that the service start time of each service object is not later than the upper bound of its time window.

[0170] 8. Service priority constraints:

[0171] If the priority of team k1 is greater than the priority of team k2, then (s k1 ir1 +t S ik1 )≤s k2 ir2 , i∈Fg, r1∈Vk1, r2∈Vk2, (12)

[0172] s k1 ir1 s represents the time when vehicle r1 in fleet k1 starts serving the target node i. k2 ir2 t represents the time when vehicle r2 in fleet k2 starts serving the target node i. S ik1 This represents the duration required for fleet k1 to provide service to service target node i. The service priority constraint ensures that, where fleet k1's service must take precedence over fleet k2's service, fleet k1's service must be completed before fleet k2's service. This service priority constraint only applies to vehicles actually assigned to service target node i.

[0173] 9. Variable type constraints:

[0174] x k ajr ∈{0,1}, a, j∈Vg, r∈Vk, k∈Q (13);

[0175] s k jr ≥0, i∈Fg, r∈Vk, k∈Q (14);

[0176] PC i CPD i CCD i LCT i ≥0, i∈Fg (15)

[0177] D i =max{0,LCT i -PBT i}, i∈Fg (16)

[0178] Among them, PC i For the meal preparation completion time of service object node i, CPD i For the cabin door closing time of service object node i, CCD i LCT is the time for closing the cargo door of service node i. i LCT is the estimated latest critical service completion time for service object node i. i =max(PC i CPD i CCD i PBT i The planned wheel removal time for service object node i. In this invention, it is assumed that each service object has only one specific fleet providing catering, passenger cabin door closing, and cargo cabin door closing services.

[0179] In this embodiment of the invention, the real-time task duration parameter set is injected as a key parameter into the optimization model in the following manner:

[0180] The set of real-time task duration parameters generated in step S320 is used as the corresponding task duration variable t in the optimization model. s ik The value is assigned accordingly. Specifically, for the service object i provided by fleet k within the current rolling time window, the task duration t in its model is... s ik It is directly assigned the task duration predicted by probability distribution quantiles in S320.

[0181] This assignment operation transforms the task duration from a random variable into a deterministic, robust parameter based on historical data statistical patterns. Essentially, it transforms the original stochastic optimization problem into a deterministic mixed-integer programming problem driven by robust input parameters, laying the mathematical foundation for generating more robust scheduling schemes.

[0182] Furthermore, the metaheuristic optimization algorithm is specifically an adaptive global optimization engine. The adaptive global optimization engine includes deconstruction operation units and reconstruction operation units. The core improvement of this engine lies in its operator selection mechanism, namely, in each iteration, the selection of deconstruction and reconstruction operation units is performed dynamically and adaptively based on a context-aware online learning module. The core of the online learning module is to formalize the operator selection problem of the adaptive global optimization engine into a sequential decision problem and solve this problem using a reward model.

[0183] The construction and workflow of the reward model are as follows:

[0184] (1) Definition of decision-making environment:

[0185] Action space: The combination of all available deconstruction and reconstruction operation units is defined as the action that the model can choose.

[0186] State space: The state of the model is defined by the state feature vector extracted from the current solution at each iteration. This feature vector quantifies the status of the current solution and includes, for example, the current total turnaround time, the proportion of vehicles with more than the average number of paths, and the difference between the latest planned start time and the current time being less than a preset threshold (i.e., (LT)). i The number of tasks (current time) < δ, where δ is a preset threshold), and the load rate of the vehicle with the highest resource utilization.

[0187] Reward signal: Whether a new solution is accepted by the algorithm after one iteration is used as an immediate reward. For example, if the new solution is accepted, it is a positive reward; otherwise, it is a negative reward or zero.

[0188] (2) Model Implementation and Integration:

[0189] This invention employs a context-based multi-armed gambling machine as the implementation framework for the reward model. Within this framework, each operator pair (action) is considered an arm, the state feature vector serves as the context for arm selection, and the reward signal is used to update the value estimate of each arm within that specific context.

[0190] Optionally, the reward model can also be implemented using a policy-based random selection model, such as the Softmax function, which randomly selects operator weights calculated based on state characteristics and updates these weights through a reward signal.

[0191] (3) Online decision-making process:

[0192] In each iteration of ALNS, perform the following operations:

[0193] The workflow of this module includes:

[0194] S331, at the beginning of each iteration, extract the state feature vector from the current solution.

[0195] S332, input the state feature vector into the reward model, and obtain the selection probability distribution output by the reward model for each available operation unit pair (i.e., the combination of deconstruction operation unit and reconstruction operation unit). The selection probability distribution reflects the expectation that using the corresponding operation unit pair in the current state can improve the solution quality.

[0196] The reward model outputs the selection probability of all operator pairs based on the policies it has learned internally (such as the expected reward of each arm in Contextual Bandit and the weight distribution in Softmax).

[0197] S333, Based on the selection probability distribution, determine the deconstruction operation unit and reconstruction operation unit used in this iteration.

[0198] In this embodiment of the invention, based on the selection probability distribution, a biased random selection strategy is used to determine the deconstruction operation unit and reconstruction operation unit used in the current iteration.

[0199] The core of this strategy is to construct a random selection function based on the selection probability of each operator pair. In a specific embodiment, this function is a roulette wheel selection function, which generates a random number in the interval [0, 1) and selects operator pairs based on the cumulative distribution function of the probability distribution, thereby ensuring that the chance of each operator pair being selected is proportional to its selection probability.

[0200] In another embodiment, the strategy may employ an ε-greedy selection function, which directly selects the current optimal operator pair with a high probability of (1-ε), while exploring completely randomly with a low probability of ε, in order to achieve a dynamic balance between exploitation and exploration.

[0201] Through the above strategy, the algorithm can dynamically adjust the search direction, both favoring operators with excellent historical performance and maintaining the ability to escape local optima.

[0202] S334, after the iteration is complete, a reward signal is generated based on whether the new solution is accepted, and this reward signal is used to update the internal policy of the reward model, thereby achieving online adaptive and continuous optimization of the operator selection policy. The reward signal is a scalar value; it is positive when the new solution is accepted, and zero or negative when the new solution is rejected. The specific value can be set hierarchically according to the degree of improvement of the new solution relative to the current solution and the historical best solution.

[0203] Specifically, using this reward signal, the internal parameters in the reward model related to the selected operator pair and the current state features are adjusted based on incremental updates.

[0204] More specifically, if the reward model is constructed based on a contextual multi-armed gambling machine, then this update process is as follows: based on the current state characteristics and the reward obtained, update the estimated value of the value function of the selected operator pair in that state.

[0205] If the reward model is built based on Softmax regression, the update process is as follows: based on the reward signal, the model weight parameters are fine-tuned using stochastic gradient descent or a variant thereof, so that higher selection probabilities are assigned to operator pairs that bring positive rewards.

[0206] In this way, the present invention precisely defines the mapping relationship between the operator selection of the adaptive global optimization engine and each element (state, action, reward) in the online learning framework, and seamlessly integrates the learning framework into the iterative loop of the adaptive global optimization engine, thus forming an efficient automated scheduling scheme generation system that can intelligently adapt to the specific problem state.

[0207] Furthermore, the deconstruction operation unit includes a delay propagation chain deconstruction operation unit. The operation of the deconstruction operation unit includes: identifying a sequence of service objects in the current scheduling scheme where subsequent tasks are at risk of cascading delays due to delays in preceding tasks, and removing some or all tasks in this sequence from the current solution. Specifically, the operation of the deconstruction operation unit includes:

[0208] Identify the source of delay: Traverse all service object nodes in the current solution. If the estimated latest critical service completion time (LCT) of service object node i is... i It was later than the planned wheel replacement time PBT i If so, then mark the service object as a source of delay.

[0209] Constructing the transmission chain: For each delay source service object node i, recursively search for its direct and indirect subsequent dependent service object set J based on the priority and dependency relationship (prior knowledge) of the guaranteed services. i Among them, the subsequent dependent service object z∈J i This means that the commencement of a certain flight operation must wait for the completion of a certain flight operation of delay source i.

[0210] Execution disruption: from the set of subsequent dependent service objects J i In the process, at least one service object z is randomly selected, and one or more of the navigation operations of the service object z are removed from its current vehicle path and placed into the unassigned task pool.

[0211] By specifically disrupting key nodes in the delay propagation path, this operator forces subsequent repair phases to replan the paths and timings for these tasks, thereby creating opportunities to break the chain reaction of delays and reconstruct feasible non-delay scheduling schemes.

[0212] Furthermore, an adaptive global optimization engine is used to solve the problem and generate service-oriented resource collaborative orchestration instructions, specifically including:

[0213] S10, Algorithm Initialization

[0214] Initial solution generation: Starting from the real-time running view state of the current rolling time window, a feasible initial scheduling scheme is generated using a fast heuristic rule (such as earliest expiration time priority, shortest processing time priority, or a combination thereof), which serves as the current solution Scurrent and the current optimal solution Sbest.

[0215] Parameter settings: Initialize algorithm parameters, including the maximum number of iterations (Iter). max Simulated annealing initial temperature T init The cooling rate α and the same initial weights are set for each deconstruction operation unit and reconstruction operation unit.

[0216] In this embodiment of the invention, the solution generated by the adaptive global optimization engine is a complete arrangement of all flight operations to be scheduled within the current rolling time window, including:

[0217] Task assignment: For each service object in Fg and each service it requires, the solution must specify which specific vehicle in which fleet will provide the service.

[0218] Path planning: For each vehicle used, the solution must define an ordered operation path. This path is a sequence of nodes that indicates which service object nodes the vehicle serves in sequence from its starting point (the garage or its current real-time location) and finally returns to the garage (or ends at the last task point). For example, the path of shuttle bus r1 is: [Garage] -> [Service object CA101] -> [Service object MU202] -> [Garage].

[0219] Scheduling: For each service object task on the path, the solution must determine a precise planned start time for service. This scheduling must satisfy all constraints defined by prior knowledge, including: service time window constraints, service priority constraints, and vehicle availability time constraints.

[0220] Objective function value: Based on the above arrangement, the solution corresponds to a quantifiable total service object turnaround delay time, which is calculated by the formula C(S)=ΣD i This value is the only objective metric for evaluating the quality of solutions generated by the adaptive global optimization engine, and it is also the target that the adaptive global optimization engine seeks to minimize. Σ represents the summation symbol.

[0221] S20, Iterative Search Process

[0222] Before the termination condition is met (such as reaching the maximum number of iterations or the solution quality not improving in consecutive iterations), repeat the following steps:

[0223] S21, Intelligent Operator Selection:

[0224] (1) Extract the state feature vector from the current solution Scurrent.

[0225] (2) Input the feature vector into the reward model to obtain the selection probability of all available deconstruction operation units and reconstruction operation units.

[0226] (3) Based on this probability distribution, the deconstruction operation unit DS and reconstruction operation unit RP used in this iteration are determined by the roulette wheel selection mechanism.

[0227] S22, Destruction Phase:

[0228] The selected deconstruction operation unit DS is applied to perturb the current solution Scurrent, removing a portion of the tasks from its vehicle paths (e.g., removing 10%-30% of the tasks), forming a partial solution Spartial and an unassigned task set Lremoved.

[0229] S23, Repair Phase:

[0230] The selected reconstruction operation unit RP is applied to attempt to re-insert tasks from the unassigned task set Lremoved into feasible positions in the partial solution Spartial, generating a new solution Snew. This process must strictly adhere to all constraints defined by prior knowledge (such as time windows, priorities, vehicle capacity, etc.).

[0231] S24, Evaluation and Acceptance of the Solution:

[0232] Calculate the objective function value (total service object turnaround time) Cnew of the new solution Snew.

[0233] The decision to accept a new solution is based on the simulated annealing criterion.

[0234] If Cnew < Ccurrent, then accept the new solution and let Scurrent = Snew. Ccurrent is the objective function value of the current solution.

[0235] If Cnew ≥ Ccurrent, then the new solution is accepted as the new current solution with probability P = exp(-(Cnew - Ccurrent) / T), where T is the current temperature. This helps the algorithm escape local optima.

[0236] If Cnew < Cbest, then update the historical best solution Sbest and set Sbest = Snew. Cbest is the objective function value of the historical best solution.

[0237] S25, Adaptive Update:

[0238] (1) Update the weights of operators DS and RP based on their performance in the current iteration (e.g., whether a new current solution or a historical best solution has been found). Operators that perform well will receive higher weights, thus increasing their probability of being selected in future iterations.

[0239] The specific update mechanism is as follows:

[0240] Performance Evaluation: Calculate a reward score σ for the operator pair used in this round. This score is based on the quality setting of the new solution Snew, for example:

[0241] If Snew is accepted as the new historical best solution Sbest, then the highest reward is given.

[0242] If Snew is accepted as the new current solution Scurrent but not the historical best solution, then a moderate reward is given.

[0243] If Snew is not accepted, zero reward or penalty will be given.

[0244] Weight Update: Using the reward score, the operator's weights w are adjusted through a smooth update formula. This formula aims to make the weight changes both responsive to recent performance and maintain a certain historical inertia, preventing excessive fluctuations. A general update form can be expressed as: w←(1-ρ)×w+ρ×σ, where ρ is a learning rate parameter between 0 and 1, for example, 0.1.

[0245] Probability synchronization: After all operators' weights are updated, they will be renormalized and converted into selection probabilities to ensure that in the next round of intelligent selection (S331), operators with excellent performance will have a higher probability of being selected.

[0246] (2) Reduce the current temperature according to the cooling plan: Ta = α × T, where Ta is the temperature after cooling and T is the current temperature.

[0247] S30, Result Output and Instruction Generation

[0248] When the algorithm terminates, it outputs the historical best solution Sbest as the final scheduling scheme for the current rolling time window.

[0249] The final scheduling plan is then translated into specific, executable service-based resource orchestration instructions. These instructions typically include a structured list of task assignments, explicitly specifying the roles for each support vehicle.

[0250] Task sequence: A list of service object numbers arranged in service order.

[0251] Planned time: The estimated start time for each task.

[0252] Travel path: The route the vehicle takes between mission points.

[0253] Through the above process, the adaptive global optimization engine can dynamically generate high-quality, executable scheduling instructions for complex, constrained multi-service resource collaborative orchestration problems within a limited computation time.

[0254] S340, Instruction output step: Output the service-oriented resource collaborative orchestration instruction.

[0255] The optimal scheduling scheme obtained from step S330 is transformed into a structured service-oriented resource collaborative orchestration instruction set that can be directly parsed and executed by ground support personnel and vehicle terminals, and then output. This scheduling instruction set is a structured data object, the core content of which includes, but is not limited to:

[0256] Global instruction header: Contains metadata of the instruction, such as the scroll window identifier, generation timestamp, scheduling cycle, and overall performance metrics (such as the estimated total turnaround time, the total number of vehicles involved, etc.).

[0257] Vehicle Task Assignment List: This is the core of the instruction set. It generates an independent task sub-instruction for each dispatched support vehicle. This sub-instruction is an ordered list that clearly defines:

[0258] Task sequence: The order in which each vehicle serves the service object nodes, for example [service object CA101, service object MU202, ...].

[0259] Planned Timeline: Specifies the planned start time and expected end time for each service object task in the sequence.

[0260] Route guidance: Specifies the recommended driving route for vehicles between mission points and can be linked to the airport's digital map system.

[0261] Task Status Update Table: Lists the latest status (such as "Assigned" or "Pending Execution") of all scheduled service objects and their various air traffic operations within the current window, as well as their corresponding service vehicles, providing a global view for the airport command system.

[0262] The specific form of the output can be:

[0263] A machine-readable data exchange format (such as JSON or XML files) is transmitted to an airport ground control system, vehicle-mounted terminal, or driver's mobile device via an application programming interface.

[0264] A visual graphical interface is presented on the dispatcher's workstation in the form of a Gantt chart or airport map path.

[0265] By outputting the structured instruction set described above, this invention achieves seamless integration from the optimization model to actual operational instructions, ensuring the automated and precise execution of the scheduling scheme.

[0266] The service-oriented resource coordination and orchestration instructions are issued to the corresponding support vehicle onboard terminals or driver mobile devices, and synchronized to the airport ground command system for monitoring and coordination.

[0267] Furthermore, the S300 also includes the following steps:

[0268] S350, Window Scrolling Step: After completing the instruction output, advance the current time by one optimization cycle Δt, then wait until a new cycle begins, automatically jump to the status acquisition step S310, and start a new round of scrolling optimization. This cycle continues until all flight operations for the service targets of the day are completed. Furthermore, the method also includes the following steps:

[0269] S400, Model and Knowledge Update Steps: Based on the deviation between the actual task execution data and the predicted data, dynamically trigger the retraining of the prediction model and / or the calibration of the prior knowledge.

[0270] Specifically, during the execution of the rolling optimization scheduling step, the actual execution time of the flight operation is collected in real time; the actual execution time is compared with the predicted time corresponding to step S320; when the deviation continues to exceed the preset tolerance, the retraining of the prediction model is triggered and / or the business rules in the prior knowledge are dynamically adjusted.

[0271] The specific implementation process of S400 is as follows:

[0272] S401, Data Acquisition and Comparison:

[0273] After each rolling window's scheduling instruction is executed, the system continuously monitors and records the actual start and end times of each flight operation, and calculates the actual execution time t based on this. actual .

[0274] t actual Compared with the predicted duration t generated in step S320 to guide this scheduling predict Perform a comparison and calculate the absolute deviation |t actual -t predict|or relative deviation|t actual -t predict | / t predict .

[0275] S402, Deviation Analysis and Trigger Judgment:

[0276] A dynamic deviation recording window is maintained for each task type. When the deviation of a certain type of task exceeds a preset tolerance threshold (e.g., relative deviation exceeds 20%) in N consecutive execution instances (e.g., 5 consecutive times), it is determined that the prediction model or related business rules for that task have become out of touch with the actual operation, triggering an update mechanism.

[0277] When the following systematic and patterned deviation characteristics are detected, it indicates that the problem may stem from outdated prior knowledge (business rules) rather than the prediction model itself:

[0278] Condition A: Group-specific systematic bias

[0279] When the bias is highly concentrated in a specific group defined by prior knowledge, while other groups perform normally.

[0280] Example: The actual duration of cabin cleaning tasks for the "A380 model" consistently and systematically exceeded the predicted value by more than 20%, while the deviation of cabin cleaning tasks for the "B737 model" remained within the normal range. This indicates that the problem likely lies in the business rule of "standard cleaning duration for the A380 model," rather than in the overall cabin cleaning prediction model.

[0281] Condition B: Changes to physical constraints or rules

[0282] When the actual duration is affected by new, known physical constraints or explicit process changes.

[0283] Example: The airport has built a remote parking stand, causing a fixed increase of 5 minutes in travel time for all air traffic operations heading to that area. This systematic bias caused by changes in spatial topology should be addressed by updating the prior knowledge of the "baseline path travel time".

[0284] When the following widespread and random bias characteristics are detected, it indicates that the problem stems from the prediction model's failure to accurately capture historical patterns:

[0285] Condition C: Pervasive bias across groupings

[0286] When the deviation is widespread across multiple or even all groups of the same task type.

[0287] Example: The actual duration of the "cabin cleaning" task, regardless of whether it's an A380, B737, or A320 aircraft, exhibits similar volatility and positive deviation. This indicates that the general "cabin cleaning duration prediction model" has become completely ineffective and needs to be retrained.

[0288] Condition D: Increased randomness bias without a clear pattern

[0289] When the variance of the bias increases significantly and cannot be attributed to any known business grouping, it indicates a decline in the model's ability to handle randomness, requiring more and newer data to learn about the uncertainty in the current environment.

[0290] S403, Model and Knowledge Update:

[0291] Predictive model retraining: Automatically add new historical data of this task type (including data before and after the trigger update) to the training set, re-execute the "predictive model building steps" of S200, and generate an updated probability distribution model that better fits the current operational patterns.

[0292] Dynamic adjustment of prior knowledge: When deviations stem from changes in business rules, the system can automatically calibrate the corresponding parameters in the prior knowledge base, or generate adjustment suggestions for administrator confirmation before updating. For example, if data indicates a systematic increase in "cabin cleaning time for A380 aircraft," it can suggest modifying the standard operating time (a type of prior knowledge) for cleaning tasks on that aircraft type. Once confirmed, this prior knowledge will be updated in subsequent scheduling.

[0293] 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.

[0294] This invention also provides a computer-readable storage medium storing computer-executable instructions for performing the methods described in this invention.

[0295] 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.

[0296] 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 interactive collaborative decision-making and autonomous scheduling method for airport ground support vehicles, characterized in that, Performed by an electronic device, including the following steps: S100, acquire and encode prior knowledge of ground support for the target airport; S200, based on the historical mission execution dataset of the target airport, builds a predictive model for mission duration for different types of air traffic operations; S300 periodically executes a rolling optimization collaborative orchestration process to generate and output service-oriented resource collaborative orchestration instructions; The S300 specifically includes: S310, obtain the real-time operational visual status at the current moment. The real-time operational visual status includes the real-time location and availability of all support vehicles, the execution status of all incomplete navigation operations, and the set of service objects that need to be served within the current rolling time window. S320, Based on the prediction model, the duration of each flight operation of each service object within the current rolling time window is predicted, and a real-time task duration parameter set is generated; the generation of the real-time task duration parameter set refers to extracting the quantile corresponding to the preset confidence level from the optimal probability distribution corresponding to the task, as the predicted duration of the task. S330, Based on the prior knowledge and the real-time running visual state, construct an optimization model with the goal of minimizing the total service object turnaround delay time within the current rolling time window; inject the real-time task duration parameter set as a key parameter into the optimization model, and use a metaheuristic optimization algorithm to solve it, generating service-oriented resource collaborative orchestration instructions; S340, output the service-oriented resource collaborative orchestration instruction; In S330, constructing an optimization model based on the prior knowledge and the real-time running visual state includes: Based on the real-time location and availability of the vehicles, the initial state is set for the path constraints of the optimization model; Update the optimization range and boundaries of the current scrolling time window based on the execution status of unfinished tasks; The constraints followed by the optimization model in S330 include: service uniqueness constraint, route continuity constraint, vehicle task triggering constraint, garage entry and exit balance constraint, time sequence and continuity constraint, vehicle capacity constraint, service time window constraint, and service priority constraint.

2. The method of claim 1, wherein, The method further includes the following steps: S400, during the execution of rolling optimization collaborative orchestration, based on the deviation between the actual task execution data and the predicted data, dynamically trigger the retraining of the prediction model and / or the calibration of the prior knowledge.

3. The method of claim 1, wherein, S200 specifically includes: Grouping historical task execution datasets based on task type and machine model; For the task duration data within each group, we perform fitting and goodness-of-fit tests on multiple candidate probability distribution models to determine the optimal probability distribution model as the prediction model for that group.

4. The method of claim 3, wherein, The candidate probability distribution model includes parametric distribution and non-parametric distribution.

5. The method of claim 1, wherein, The metaheuristic optimization algorithm is an adaptive global optimization engine; wherein, the adaptive global optimization engine includes a deconstruction operation unit and a reconstruction operation unit, and the selection of the deconstruction operation unit and the reconstruction operation unit is implemented through a context-aware online learning module.

6. The method of claim 5, wherein, The workflow of the online learning module includes: S331, at the beginning of each iteration, extract a state feature vector from the current solution. The state feature vector includes at least the current total turnaround delay time, the number of tasks whose difference between the planned latest start time and the current time is less than a preset threshold, and the load rate of the vehicle with the highest resource utilization. S332, Input the state feature vector into a reward model, and obtain the selection probability distribution of the output for each available operation unit in the reward model. The selection probability distribution reflects the expectation that using the corresponding operation unit in the current state can improve the solution quality. S333, Based on the selection probability distribution, determine the deconstruction operation unit and reconstruction operation unit used in this iteration; S334, after the iteration is completed, a reward signal is generated based on whether the new solution is accepted, and the internal policy of the reward model is updated using the reward signal.

Images