Dynamic optimization method and system for airport ground support resources based on digital twinning

Abstract

The invention relates to the technical field of airport resource management, in particular to an airport ground support resource dynamic optimization method and system based on digital twinning, and the method comprises the steps: analyzing an airport business process knowledge graph, and fusing multi-modal data to construct a generative digital twinning model; performing multi-scene simulation deduction and multi-target optimization based on the generative digital twin model to generate a long-term planning scheme; during the execution period of the long-term planning scheme, monitoring the deviation between the actual operation data and the scheme predicted value, and performing dynamic adaptive adjustment on the long-term planning scheme; and combining a long-term planning scheme with real-time operation data to provide decision support, evaluating an optimization effect, and taking an evaluation result as feedback to iteratively optimize the generative digital twinborn model. According to the method, the generation type digital twinborn model is constructed, and a double-layer closed-loop optimization mechanism of long-term planning and short-term correction is established, so that the overall efficiency and punctuality rate of airport ground guarantee and the toughness of coping with emergencies can be effectively improved.

Description

Technical Field

[0001] This invention relates to the field of airport resource management technology, and more specifically, to a method and system for dynamic optimization of airport ground support resources based on digital twins. Background Technology

[0002] With the rapid development of the global aviation industry, the number of flights and passenger throughput at major hub airports continues to rise, leading to an increasing workload and complexity for airport ground support systems. Ground support services, as a crucial link in ensuring flight punctuality and enhancing passenger experience, encompass a series of tightly coupled operational processes, from aircraft parking, passenger bridge docking, baggage handling, and catering to cleaning, refueling, and de-icing. This series of processes involves the coordinated scheduling of multiple types of large-scale resources, including personnel, vehicles, equipment, and berths, and its scheduling efficiency directly determines the overall operational efficiency of the airport. In actual operation, unpredictable events such as sudden weather changes, flight delays, and equipment failures occur frequently, posing a severe challenge to traditional resource scheduling models. To cope with this highly dynamic and complex operating environment, the industry has been exploring the use of advanced digital and intelligent technologies to manage and optimize airport ground support resources more precisely, intelligently, and proactively. Currently, the limitations of existing airport resource management technologies are becoming increasingly apparent when dealing with highly dynamic and complex operating environments. On the one hand, many systems rely on fixed rules and static scheduling schemes based on historical experience, lacking a deep understanding of the complex and dynamic relationships between various support links in an airport. This results in resource allocation often being locally optimal rather than globally optimal, making it difficult to adapt to rapidly changing actual operating conditions. On the other hand, existing airport resource management technologies are insufficient in utilizing the massive amounts of multi-source heterogeneous data generated during airport operations. In particular, there is a lack of effective automatic parsing and utilization methods for unstructured text data containing rich business logic, such as operation manuals and emergency plans, leading to a disconnect between digital models and real business processes. Summary of the Invention

[0003] The purpose of this invention is to provide a method and system for dynamic optimization of airport ground support resources based on digital twins. By constructing a generative digital twin model and establishing a two-layer closed-loop optimization mechanism of long-term planning and short-term correction, it achieves dynamic, forward-looking and adaptive optimization of airport resources, which can effectively improve the overall efficiency, punctuality rate and resilience of airport ground support in response to emergencies.

[0004] This invention is achieved through the following technical solution: A method for dynamic optimization of airport ground support resources based on digital twins, the steps of which include: The knowledge graph of airport business processes is analyzed and multimodal data is integrated to construct a generative digital twin model that reproduces and extrapolates the dynamic operation of airport ground support. Based on the generative digital twin model, multi-scenario simulation and multi-objective optimization are performed on multi-modal data to generate a long-term planning scheme covering a set future time. During the execution of the long-term planning scheme, the deviation between the actual operating data and the predicted values ​​of the scheme is monitored in real time. When the deviation is within a preset threshold, the parameters of the generative digital twin model are adaptively corrected. When the deviation exceeds the preset threshold, the long-term planning scheme is replanned to achieve dynamic adaptive adjustment of the long-term planning scheme. By combining the long-term planning scheme with real-time operational data, the system provides managers with visual decision support, evaluates the optimization effect, and uses the evaluation results as feedback to iteratively optimize the generative digital twin model.

[0005] Optionally, a generative digital twin model is constructed, specifically by building an airport business process knowledge graph and an intelligent agent simulation model to accurately map airport ground support scenarios. The airport business process knowledge graph is calculated using the following formula:

[0006] Where G represents the business process knowledge graph; V is the set of nodes, including business entity nodes. Activity Nodes and rule nodes E is the set of edges, representing the relationships between nodes; L is the set of attribute labels for nodes and edges. Each agent in the agent simulation model The internal structure of is calculated using the following formula:

[0007] in, A unique identifier for an intelligent agent; The type of intelligent agent; At the agent level; The state space of the agent; The action space of the intelligent agent; The policy function of the agent; Let be the characteristic function of the agent.

[0008] Optionally, the intelligent agent simulation model defines the following reward function, the calculation formula of which is:

[0009] in, For agent i to take action a in state s to transition to a new state The reward value obtained at that time; Rewards for completing the task; Rewards for collaboration; As a reward for overall system performance; To balance the weighting coefficients of different rewards; The strategy of the agent simulation model is iteratively optimized using the Q-value update formula, specifically as follows:

[0010] in, The state-action value of agent i taking action a in state s; The learning rate; Discount factor; In the new state The maximum Q value among all possible actions.

[0011] Optionally, the multimodal data is specifically integrated by constructing a multimodal data fusion framework to integrate multi-source heterogeneous data. This multimodal data fusion framework utilizes a self-attention mechanism to extract and fuse multimodal data features, and its calculation formula is as follows:

[0012] Where Q is the query matrix; K is the key matrix; and V is the value matrix; The dimension of the key vector; Based on the reliability assessment results of each data source, the fused feature vector is dynamically weighted and updated using the following formula:

[0013] in, for The fused feature vector at each time step; For the k-th data mode in The feature vector at time step; Let be the dynamic fusion weight for the k-th data modality, and .

[0014] Optionally, the multi-objective optimization defines three objectives: maximizing resource utilization, minimizing operating costs, and maximizing service quality. These objectives are then transformed into a single-objective optimization problem through the following comprehensive optimization objective function:

[0015] in, Let X be the overall objective function to be maximized; X be the resource allocation scheme. These are the weighting coefficients for each objective; The objective function is to maximize resource utilization. The objective function is to minimize operating costs; The objective function is to maximize service quality. Meanwhile, the solution process satisfies resource availability constraints, task requirement constraints, and resource transfer constraints.

[0016] Optionally, the adaptive correction of the parameters of the generative digital twin model specifically involves constructing a short-term dynamic correction optimization problem, which includes minimizing the deviation from long-term planning and maximizing the benefits of short-term optimization. The optimization objective of minimizing the deviation from the long-term plan is calculated using the following formula:

[0017] in, This is a short-term dynamic correction scheme; This is a long-term planning scheme; and In the short-term and long-term scenarios, respectively, in terms of time... The amount of resource i allocated to task j; To optimize the time window in the short term; The calculation formula for the optimization objective of maximizing the benefits of short-term optimization is as follows:

[0018] in, To improve resource utilization rate; Cost reduction rate; To improve service quality rate; These are the weighting coefficients for each benefit indicator.

[0019] Optionally, the specific calculation formula for evaluating the optimization effect is as follows:

[0020] in, A comprehensive score for the decision-making options; to These are the weighting coefficients for each evaluation indicator; As a relative evaluation indicator for resource utilization efficiency; This is an indicator of relative changes in operating costs; For the relative change in service quality; This is an indicator of the relative change in response time. It serves as a robustness evaluation metric for the system.

[0021] A digital twin-based dynamic optimization system for airport ground support resources includes: The model building module parses the airport business process knowledge graph and integrates multimodal data to construct a generative digital twin model that reproduces and extrapolates the dynamic operation of airport ground support. The long-term planning module, based on the generative digital twin model, performs multi-scenario simulation and multi-objective optimization on multimodal data to generate a long-term planning scheme covering a set future time. The dynamic correction module monitors the deviation between the actual operating data and the predicted values ​​of the plan in real time during the execution of the long-term planning scheme. When the deviation is within a preset threshold, it adaptively corrects the parameters of the generative digital twin model. When the deviation exceeds the preset threshold, it triggers the replanning of the long-term planning scheme to achieve dynamic adaptive adjustment of the long-term planning scheme. The decision-making and feedback module combines the long-term planning scheme with real-time operational data to provide managers with visual decision support, evaluate the optimization effect, and use the evaluation results as feedback to iteratively optimize the generative digital twin model.

[0022] Optionally, the model building module also defines a compliance penalty mechanism when parsing the airport business process knowledge graph, and its calculation formula is as follows:

[0023] in, For the penalty function; The model builds a knowledge-enhanced generative model in the model building module; c represents the input context. This is the embedding vector for domain knowledge; R represents the model parameters; R is the set of domain rules. This is an indicator function, which is 1 when the condition is met and 0 otherwise; Let be the penalty coefficient for rule r.

[0024] Optionally, the dynamic correction module adaptively adjusts the weight coefficients in the optimization objective of maximizing short-term optimization benefits using the following formula:

[0025] in, These are the weights of resource utilization rate, cost reduction rate, and service quality improvement rate at time t, respectively. The learning rate; Let be the overall benefit function at time t; These are the gradients of the overall benefit function with respect to each weight coefficient.

[0026] The technical solution of the present invention has at least the following advantages and beneficial effects: This invention constructs a generative digital twin model that deeply integrates real-time multimodal data and unstructured business process knowledge from airports, creating a living digital mirror of the airport that accurately reproduces the real world and can predict the future. Furthermore, this invention innovatively designs a two-layer closed-loop optimization mechanism combining long-term planning and short-term adjustments. This mechanism not only generates forward-looking long-term resource optimization solutions based on multi-scenario simulations for proactive and preventative management, but also monitors actual operational deviations in real time. Through parameter adaptive correction and dynamic replanning, it flexibly responds to various emergencies, ensuring the continued effectiveness of the solutions and the resilience of the system. Simultaneously, the feedback optimization loop of this invention enables the entire system to self-evolve, continuously iterating and optimizing its digital twin model and decision-making capabilities based on actual operational results, ultimately achieving continuous and adaptive improvement in airport ground support efficiency. Attached Figure Description

[0027] Figure 1 A flowchart illustrating the dynamic optimization method for airport ground support resources based on digital twins provided by this invention; Figure 2 This is a schematic diagram illustrating the principle of the dynamic optimization system for airport ground support resources based on digital twins provided by the present invention. Detailed Implementation

[0028] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0029] Reference Figure 1 , Figure 2 As shown, this embodiment provides a method for dynamic optimization of airport ground support resources based on digital twins and a system for implementing this method. The system can be physically deployed in a data center server cluster within the airport or provided through a cloud computing platform. The entire system is divided into four closely cooperating and logically progressive core functional modules: a model building module, a long-term planning module, a dynamic correction module, and a decision-making and feedback module. These four modules together constitute a two-layer closed-loop intelligent optimization system, capable of proactive, dynamic, and adaptive optimization and scheduling of airport ground support resources to cope with complex and ever-changing operating environments.

[0030] In this specific implementation, the model building module is the foundation of the entire optimization system. Its core task is to build a generative digital twin model that can accurately map the physical world of the airport, deeply understand business processes, and extrapolate various future possibilities. The construction process of this module integrates AI-driven business process analysis, multi-agent simulation modeling, multi-modal data fusion, and knowledge-enhanced generative AI technologies. Specific implementation details are as follows: First, this module constructs a business process knowledge graph. It automatically parses a large amount of unstructured and semi-structured text data relied upon by airport operations, such as the Airport Operations Manual, Emergency Response Plan, and Standard Operating Procedures, using advanced natural language processing technology. Through named entity recognition technology, it accurately identifies key business entities in the text, such as flights, passengers, boarding gates, jet bridges, and baggage carousels. Through relation extraction technology, it parses the static subordinate or associated relationships between entities, such as which terminal a particular boarding gate belongs to. Through event extraction technology, it identifies key business activities, such as check-in, security checks, boarding, and baggage handling, and captures the complex logical and temporal dependencies between these activities.

[0031] All extracted knowledge elements are organized into a structured business process knowledge graph, the overall structure of which is defined by formula (1):

[0032] Where G represents the business process knowledge graph; V is the set of nodes, including business entity nodes. It encompasses all physical or logical entities and activity nodes within the airport. This represents the various specific operational steps and rule nodes of ground support. E is used to encapsulate various constraints in the assurance process, such as security regulations, priority rules, and time window limits; E is the set of edges, representing the relationships between nodes, i.e. ,in, This represents the execution or participation relationship between an entity and an activity. This indicates the sequence or prerequisite dependencies between different activities. This indicates that the activity is subject to specific rules, and This indicates that the rule applies to a specific entity; L is a set of attribute labels for nodes and edges, used to store detailed information for each element. For example, the attributes of a flight node may include flight number, scheduled departure time, aircraft type, etc., while the attributes of an edge can indicate the type of relationship, such as "prerequisite task" or "resource requirement".

[0033] To optimize the structure of the knowledge graph and make it better reflect core business processes, this module calculates and quantifies the strength of relationships between entities and activities. This relationship strength measures the closeness of the association between business entities and specific support activities, and its calculation follows formula (2):

[0034] in, This represents the numerical value of the relationship between entity e (e.g., a specific flight) and activity a (e.g., the check-in service for that flight). The calculation of this value combines two core metrics: one is... This represents the frequency with which entity e and activity a appear simultaneously in historical operational data; high frequency usually indicates a strong correlation. Secondly... , representing the semantic relevance score between entity e and activity a, calculated using an advanced semantic analysis model. and These are two adjustable weighting coefficients that sum to 1, used to balance the contributions of historical co-occurrence frequency and current semantic relevance in the final strength calculation. By sorting the relation strength values ​​and setting thresholds for filtering, a more refined, efficient, and practical business process knowledge graph can be constructed.

[0035] To ensure the knowledge graph can keep pace with the times, this module also establishes a dynamic update mechanism. This mechanism can monitor and receive various new business documents and rule change notifications issued by the airport information system in real time. The dynamic update process of the knowledge graph is mathematically described by formula (3):

[0036] in, This represents the complete state of the knowledge graph at time t. Indicates from time t to During this specific timeframe, knowledge content needs to be analyzed and added to the graph, such as new support procedures required due to the introduction of new passenger aircraft at the airport, or newly released emergency response plans for extreme weather. Accordingly, This represents outdated or obsolete knowledge content that needs to be removed from the graph during this period, such as invalid operational procedures. Through this incremental "metabolism" mechanism, the knowledge graph can always maintain the freshness and accuracy of its content.

[0037] Secondly, guided by the constructed business process knowledge graph, this module begins to create a highly realistic intelligent agent simulation model. The goal of this model is to abstract all entities with proactive behavior capabilities in the airport ground support scenario, such as flights, passengers, ground staff, support vehicles, baggage systems, etc., into intelligent agents with independent attributes, states, and behavioral decision-making rules. The set of these intelligent agents is defined by formula (4):

[0038] in, Each of these represents an independent intelligent agent in the system. Each agent is able to perceive the state of its environment and make autonomous decisions based on its internal goals and policy functions, thereby simulating highly realistic complex behaviors and interactions in the simulated world.

[0039] To accurately describe and define each agent, its internal structure is characterized in detail by the tuples defined by formula (5):

[0040] in, A unique identifier for the intelligent agent, such as flight number "MU5101" or device number "TUG-021"; For intelligent agent types, such as flights, passengers, ground crew, refueling trucks, and jet bridges; The agent hierarchy is used to represent the importance or hierarchical relationship of the agents in the system. For example, the hierarchy of the flight agent is usually higher than that of a single support vehicle agent. The state space of an agent is a set that describes all its possible states. For example, the states of a flight agent may include "planned", "arrived", "berthed", "under maintenance", "delayed", "departed", etc. The action space of an agent is also a set that describes all the actions it can perform. For example, a flight agent can perform actions such as "requesting a gate", "requesting de-icing service", and "reporting delays". The policy function is the core of the agent's intelligent decision-making, defining the rules and probabilities of which action to choose in a specific state; is the feature function of the agent, used to extract the key features most valuable for decision-making from its complex raw state information.

[0041] The simulation model operates using a discrete event-driven mechanism to simulate the dynamic evolution of the entire airport system. The transition of the system's global state is driven by the state transition function defined by equation (6):

[0042] here, It represents the global state of the entire system at time t, that is, a snapshot of the current state of all agents. This represents the set of actions taken by all agents according to their respective policies at time t. This represents the set of external random events occurring at time t, such as a sudden thunderstorm, a temporary lockdown of a region, or an unexpected failure of critical equipment. T is the core state transition function, which accurately calculates the next simulation time point based on the current system state, the collective actions of all agents, and the unexpected events in the external environment. The new state of the system drives the simulation process forward step by step.

[0043] The time progression mechanism of the simulation model follows the rules defined by formula (7):

[0044] in, It is an event queue in the current simulation world that is sorted by time and stores all planned or triggered events that have not yet been processed. It is the planned occurrence time of each event e. The next time step of the simulation is determined by taking the earliest occurrence time among all events to be processed. Through this mechanism, the simulation model can accurately simulate and handle various complex concurrent events according to the actual occurrence order of events, ensuring the accuracy of the timing logic throughout the simulation process.

[0045] To guide agents to learn optimal behavioral strategies, rather than simply passively executing preset rules, this module introduces a reinforcement learning algorithm. A refined reward function is designed for each agent to incentivize them to make decisions beneficial to the overall system goal. The reward function for agent i is defined by formula (8):

[0046] in, For agent i to take action a in state s to transition to a new state The reward value obtained at that time; Rewards for completing the task; Rewards for collaboration; As a reward for overall system performance; To balance the weighting coefficients of different rewards, this formula calculates the weighting coefficients for actions taken by agent i in state s that result in a transition to a new state. The total reward value that can be obtained at that time. This reward is composed of three weighted parts: the first part is This refers to the task completion reward; for example, completing a task on time will earn a positive reward, while failing to complete it on time will result in a negative reward penalty. The second part is... This refers to collaboration rewards. When an agent's actions directly help other agents complete their tasks more efficiently, it receives a reward, thus encouraging team collaboration among agents. The third part is... This refers to the system-wide performance reward. If the agent's behavior helps improve the airport's overall operational indicators, such as the overall flight punctuality rate or resource utilization rate, it will also receive a corresponding reward. These are the weighting coefficients of these three parts of the reward, and their sum is 1. They can be dynamically adjusted according to the airport's operational goals at different times (e.g., prioritizing on-time performance or prioritizing cost reduction).

[0047] The agent continuously updates its internal policy function by engaging in trial-and-error interactions in a simulation environment and based on the rewards obtained from its actions. The iterative optimization process of the policy is implemented using the classic Q-learning algorithm, the core of which is shown in formula (9):

[0048] in, The state-action value of agent i taking action a in state s; The learning rate controls the magnitude by which the Q-value is updated each time based on new experience; The discount factor represents the importance of future rewards relative to current rewards; the closer its value is to 1, the more "foresighted" the agent is. In the new state The maximum Q-value among all possible actions. Through thousands of iterations of this update formula, the agent can gradually learn which action to choose in any state to maximize its future long-term total reward, thus forming the optimal decision-making strategy.

[0049] Furthermore, to provide rich and real-time data input for the simulation model, this module is responsible for building a powerful multimodal data fusion framework. This framework first connects to various heterogeneous data sources involved in airport operations by defining unified data interface standards and protocols. The set of these data sources is defined by formula (10):

[0050] in, These represent data streams from different systems with different data modes. For example, It is flight status message data (structured numerical type) from the air traffic control system. These are weather forecasts (text and images) issued by meteorological departments in both text and radar image formats. It is time-series data (time series type) of vehicle and equipment status and location from the resource management information system. It is passenger queue length data (video stream analysis results) obtained from the analysis of surveillance cameras inside the terminal. These are the text data such as business rules and operation manuals used when constructing the knowledge graph.

[0051] The collected raw data first undergoes a series of preprocessing steps, including data cleaning to remove noise and outliers, data format conversion to standardize data, and time alignment across data sources to ensure all data are on the same time base. Subsequently, this module constructs a deep feature extraction and fusion model based on a self-attention mechanism to deeply mine the potential, non-linear intrinsic relationships between different data modalities. The core computational process of the self-attention mechanism is defined by formula (11):

[0052] Where Q is the query matrix, representing the feature representation of the information that needs to be focused on; K is the key matrix, representing the representation of all queryable feature information in the data; and V is the value matrix, representing the actual numerical content of these features. This represents the dimension of the key vector. This is a core computation function used to achieve dynamic weighted fusion of multimodal data features. By calculating the dot product of the query matrix Q and the key matrix K, the model obtains the relevance scores between different features, which are then normalized using the softmax function to obtain a set of attention weights. Finally, these weights are applied to the value matrix V, resulting in a novel feature representation that is dynamically weighted and incorporates global contextual information. The denominator contains... It is the dimension of the key vector, used to scale the dot product result to prevent the gradient from being too small during training.

[0053] Through this powerful self-attention mechanism, feature vectors from different data modalities are efficiently fused. The fused feature vector is represented by formula (12):

[0054] here, These represent different data modalities. The high-dimensional feature vectors are extracted by their respective feature extractors (e.g., CNN for images, LSTM for time series). Through self-attention computation, the model can automatically learn which modalities are more important in the current context, which features have stronger correlations, and generate a more comprehensive and informative fused feature vector than any single modality. This fused feature vector will serve as the primary input for the digital twin model to perceive the environment, providing richer and more accurate information support for subsequent simulation and decision optimization.

[0055] In addition, this module establishes a real-time data quality assessment mechanism to continuously monitor the reliability of each data source. The reliability assessment score of data source i at time t is calculated using a dynamic function of formula (13):

[0056] This function is an exponentially smoothed dynamic evaluation function, where, It is the overall reliability score of data source i at time t. Its calculation combines two parts of information: It is its historical reliability value at the previous moment, and It is based on the data received at the current moment. The instant score is calculated based on quality indicators (such as data completeness, timeliness, consistency, accuracy, etc.). It is a historical weighting coefficient between 0 and 1, used to balance the stability of historical evaluation results and the sensitivity of current evaluation results.

[0057] The reliability assessment results of the data source will be directly used to guide the fusion process of multimodal data. The data fusion adopts a dynamic weighted update mechanism, as shown in formula (14):

[0058] in, for The fused feature vector at each time step; For the k-th data mode in The feature vector at time step; Let be the dynamic fusion weight for the k-th data modality, and .

[0059] Finally, to ensure the feasibility and compliance of the various solutions and suggestions generated by this module, this module deeply integrates generative AI technology with domain knowledge of airport ground support. First, the massive amounts of structured knowledge, such as business rules, operational specifications, and expert experience, from the aforementioned business process knowledge graph are vectorized and embedded using a deep neural network. The knowledge embedding process is completed by the function defined in formula (15):

[0060] This is a domain knowledge embedding function, where K represents the input structured set of domain knowledge (e.g., subgraphs or rule paths extracted from a knowledge graph). It is a specially designed knowledge embedding neural network model, such as TransE, TransH, or GraphSAGE. The core function of this model is to map discrete, symbolic domain knowledge K into a low-dimensional, continuous vector representation. This knowledge embedding vector This enables neural networks to "understand" and utilize this complex domain knowledge.

[0061] Then, this module constructs a knowledge-enhanced generative model based on the advanced Transformer architecture. This model innovatively adds a knowledge enhancement module to the standard encoder-decoder structure. The structure of this model can be abstractly represented by equation (16):

[0062] in, This is the encoder part, which is responsible for processing the contextual information c of the input (such as a comprehensive description of the current airport state, i.e., fusing feature vectors). Encoding is performed to extract its deep semantic features. The key innovation lies in the intermediate attention module, which calculates the features output by the encoder and the aforementioned knowledge embedding vector. Attention weights are assigned between features to enable dynamic interaction and deep fusion of domain knowledge with current context features. The knowledge-enhanced feature vectors are then fed into the decoder. Ultimately, it generates a textual description of an optimization scheme, prediction result, or decision suggestion that conforms to grammar and logic. This represents all trainable parameters of the entire generative model.

[0063] When training this generative model, the system's optimization objective is twofold: not only must the accuracy of the generated content be considered, but it must also be strictly ensured that it complies with the airport's business rules. Therefore, the overall optimization objective function combines the cross-entropy loss used to evaluate prediction accuracy with the business rule compliance penalty used to ensure compliance. The cross-entropy loss function is defined by formula (17):

[0064] in, It is the output probability distribution predicted by the model. This is the actual label value (i.e., the standard answer). The loss function measures the difference between the model's prediction and the actual result; the smaller the difference, the smaller the loss.

[0065] The business rule compliance penalty mechanism is implemented through the penalty function defined by formula (18):

[0066] in, For the penalty function; The model builds a knowledge-enhanced generative model in the model building module; c represents the input context. This is the embedding vector for domain knowledge; R represents the model parameters; R is the set of domain rules. This is an indicator function, which is 1 when the condition is met and 0 otherwise; Let be the penalty coefficient for rule r.

[0067] Ultimately, the optimization objective of the model parameters is to minimize the sum of accuracy loss and rule penalty, as shown in equation (19):

[0068] By employing optimization algorithms such as gradient descent to solve this joint optimization problem, a set of optimal model parameters can be obtained. These parameters enable the model to ensure the accuracy of the generated content while strictly adhering to the airport's various business rules, thereby producing practical, safe, and compliant output.

[0069] To enable the model to adapt to continuous business changes and achieve continuous online learning, this module also designs a dynamic update rule for knowledge embedding, as shown in formula (20):

[0070] in, It is the knowledge embedding vector at time t. It is at t to New or updated domain knowledge during this period. It is an update coefficient (learning rate) used to control the speed and weight at which new knowledge is integrated into the existing knowledge system. Through this incremental online update, the model can continuously and automatically absorb new business rules and operational experience, maintaining the advanced nature and timeliness of its decision-making capabilities.

[0071] At this point, a generative digital twin model capable of accurately mapping the physical world, dynamically integrating multi-source data, deeply understanding business processes, and generating compliant intelligent solutions has been completed. This model provides a solid and reliable foundation for subsequent modules such as long-term planning, short-term adjustments, and decision support.

[0072] In this embodiment, the long-term planning module utilizes the generative digital twin environment established by the model building module to formulate a globally optimal and forward-looking ground support resource allocation and scheduling scheme for a relatively long time period (e.g., the next 24 hours, 48 ​​hours, or a week). This scheme is the core output of the system's outer optimization loop.

[0073] First, this module defines the time frame for long-term planning. Then, it invokes a digital twin model, inputting predictions of future weather, flight schedules, and other information, and uses large-scale simulations to generate time series forecasts of demand for various ground support tasks within a future timeframe. At the same time, the system will comprehensively inventory and assess the available quantity of various existing support resources. Fixed costs and operating costs that vary with usage .

[0074] This module constructs a multi-objective optimization problem, aiming to simultaneously achieve multiple interrelated or even conflicting operational objectives. In this embodiment, it mainly includes three core objectives: maximizing resource utilization, minimizing total operating costs, and maximizing service quality. These three objectives are mathematically defined by formulas (21), (22), and (23), respectively: Objective function for maximizing resource utilization :

[0075] Objective function for minimizing operating costs :

[0076] Objective function for maximizing service quality (Based on flight punctuality rate):

[0077] In these formulas, The core decision variable is denoted by , representing the quantity of resource type i allocated to task j at time point t. X is a matrix composed of all decision variables.

[0078] The optimization process must be carried out under the premise of satisfying a series of real-world constraints. These constraints mainly include the resource availability constraint defined by formula (24) (the resources allocated at any time cannot exceed the total available amount), the task requirement constraint defined by formula (25) (the resources allocated to the task must meet the minimum requirements of the task), and the resource transfer constraint defined by formula (26) (considering that it takes time for resources to move between different locations).

[0079] Since directly solving multi-objective problems is very complex, this module adopts a weighted summation method to transform the above multi-objective problem into a single-objective optimization problem, as shown in formula (27):

[0080] in, Let be the overall objective function to be maximized; The weights of each objective represent the relative importance that airport management places on the three objectives: resource utilization, cost control, and service quality. Subsequently, this module employs an improved genetic algorithm or other efficient heuristic search algorithms (such as particle swarm optimization, simulated annealing, etc.) to solve this complex combinatorial optimization problem, ultimately generating an optimal long-term planning scheme. The plan details which resources should be used to execute each support task in each time slot during the future planning period. This plan will then be distributed to the airport's operations execution system and serve as a baseline and guiding framework for the next level of short-term corrective loop.

[0081] In this embodiment, the dynamic correction module constitutes the real-time correction loop in the inner layer of the system. Its core function is to continuously monitor the deviation between the actual operating status and the plan during the execution of the long-term planning scheme, and to make rapid, local dynamic adjustments and adaptive optimizations based on the changes that occur in real time.

[0082] This module operates within a relatively short rolling time window. It operates within this window. Through the data fusion framework of the model building module, it collects real-time operational data within that window. For example, actual flight arrival times, reports of sudden equipment malfunctions, and abnormal surges in passenger traffic. Then, this real-world data is integrated with long-term planning schemes. The planned values ​​for the corresponding time period are precisely compared, and the deviation between the two is calculated.

[0083] If the calculated deviation is within the preset reasonable threshold range, it indicates that the actual situation is not significantly different from the plan. At this time, this module will initiate a short-term correction process with adaptive parameters. It will construct a brand-new, smaller-scale short-term dynamic correction optimization problem. This problem has two core optimization objectives: one is to minimize the deviation from the original long-term planning scheme under the premise of satisfying the current real-time constraints, so as to ensure the stability and continuity of the plan. Its objective function is defined by formula (28):

[0084] in, This is a short-term dynamic correction scheme; This is a long-term planning scheme; and In the short-term and long-term scenarios, respectively, in terms of time... The amount of resource i allocated to task j; To optimize the time window in the short term.

[0085] Secondly, based on this, the immediate benefits that short-term adjustments can bring are maximized, such as eliminating delays as soon as possible and reducing resource conflicts in the current period. The objective function is defined by formula (29):

[0086] in, To improve resource utilization rate; Cost reduction rate; To improve service quality rate; These are the weighting coefficients for each benefit indicator.

[0087] Solving this short-term optimization problem also requires satisfying a series of real-time constraints, such as the deviation magnitude constraint defined by formula (30). This formula defines the deviation constraint from the long-term plan, which is used to control the degree of deviation between the short-term dynamic adjustment scheme and the long-term plan. It requires that for all resources i, tasks j, and time... The deviation in resource allocation between the short-term revision plan and the long-term plan must not exceed the maximum allowable adjustment amount. This constraint ensures that the magnitude of short-term dynamic adjustments is within a reasonable range, avoiding operational chaos and resource waste caused by excessive adjustments, while maintaining basic consistency with long-term planning. The actual resource availability constraint at the current moment is defined by formula (31). This formula defines the real-time resource availability constraints, which are used to ensure that short-term dynamic adjustment schemes conform to the actual resource state. The requirements are as follows: for all resources i and time... In the short-term adjustment plan, the sum of the resources i allocated to various tasks j must not exceed the amount of resources i in time. Actual available quantity This constraint takes into account real-time changes in resource status (such as equipment failure, staff absence, etc.) to ensure the feasibility and executability of short-term dynamic correction schemes. It also considers the actual task requirements constraint at the current moment as defined by formula (32). This formula defines the constraints for real-time task requirements, serving as constraints to ensure that short-term dynamic correction schemes meet real-time task requirements. It requires that for all tasks j and time... The sum of the effective production capacity of various resources i allocated to task j in the short-term adjustment plan The time limit must not be lower than that of task j. Actual needs This constraint takes into account real-time changes in task requirements (such as flight delays and passenger flow fluctuations), ensuring that short-term dynamic correction schemes can effectively meet the actual operational needs of airport ground support and guarantee the smooth operation of core businesses such as flight operations and passenger services. This module uses Model Predictive Control (MPC) or other fast online optimization algorithms to efficiently solve this problem, thereby generating an optimal short-term correction action or scheme. This is used to replace or adjust existing long-term plans.

[0088] More importantly, this module possesses adaptive learning capabilities. After performing a short-term correction, it evaluates the actual effect of the correction and, based on the effectiveness, adjusts the weight coefficients in the short-term optimization objective function (Equation 29) using gradient descent. Adaptive fine-tuning is performed, as shown in formulas (33), (34), and (35):

[0089] in, These are the weights of resource utilization rate, cost reduction rate, and service quality improvement rate at time t, respectively. The learning rate; Let be the overall benefit function at time t; These are the gradients of the overall benefit function with respect to each weight coefficient.

[0090] This means that if a certain instance prioritizes "time" ( The significant corrections have yielded excellent results, meaning that in similar scenarios in the future, the system will be more inclined to prioritize time. This mechanism allows the system to learn from each actual operation and intervention, continuously optimizing its short-term decision preferences and model parameters, making it increasingly adaptable to the operational rhythm and characteristics of a specific airport.

[0091] Conversely, if the detected deviation exceeds a preset critical threshold, such as in the event of widespread flight delays or extreme weather warnings, this module determines that the current long-term planning scheme is no longer applicable. In this case, it will not perform local corrections but will immediately trigger a "replanning" mechanism, issuing an instruction to the long-term planning module, requiring it to re-run and generate a completely new long-term planning scheme based on entirely new inputs containing information about the critical events.

[0092] In this embodiment, the decision-making and feedback module is the top layer of the entire optimization system. It is both the human-computer interaction interface for airport management personnel and a key link in realizing the learning, evolution, and closed-loop improvement of the entire system.

[0093] This module provides comprehensive decision support for decision-makers. It clearly displays resource scheduling plans generated by the long-term planning and dynamic adjustment modules through highly visualized dashboards, Gantt charts, and 3D scenes. More importantly, it can use generative digital twin models to perform "What-if" scenarios, predicting how key performance indicators (KPIs) of the airport, such as flight punctuality, jet bridge utilization, and passenger waiting times, will change after adopting a particular solution. When the system encounters complex decision points requiring human intervention, it automatically generates multiple candidate solutions, each with its own advantages and disadvantages. And a detailed, multi-dimensional quantitative evaluation of each solution was conducted.

[0094] The evaluation indicator system is very comprehensive, aiming to provide decision-makers with a 360-degree decision-making perspective. These indicators include the relative evaluation indicators of resource utilization efficiency defined by formula (36). :

[0095] The relative change index of operating costs defined by formula (37) :

[0096] The relative change index of service quality defined by formula (38) :

[0097] The relative change index of emergency response time defined by formula (39) :

[0098] and the system robustness (i.e. anti-interference capability) evaluation index defined by formula (40) :

[0099] To provide an intuitive total score and ranking suggestion, this module uses a configurable multi-dimensional comprehensive evaluation function to score each candidate solution, as shown in formula (41):

[0100] in, A comprehensive score for the decision-making options; to The weighting coefficients for each evaluation indicator can be set by managers based on current management priorities. The highest-scoring solution will be recommended to decision-makers by the system as the preferred option. As a relative evaluation indicator for resource utilization efficiency; This is an indicator of relative changes in operating costs; For the relative change in service quality; This is an indicator of the relative change in response time. It serves as a robustness evaluation metric for the system.

[0101] This module is responsible for closing the learning loop of the entire system. After any decision-making scheme (whether automatically executed or manually selected) is implemented, this module continuously tracks and collects the operational data and final results of that scheme in the real world. Then, it calculates the true values ​​of various evaluation indicators and compares these true values ​​with the predicted values ​​before the scheme was implemented, generating a detailed "prediction-execution" effect evaluation report. This report not only helps managers review the effectiveness of decisions, but more importantly, it contains valuable empirical data. This module structures these validated decision cases, deviation analyses, evaluation results, and other information, and then feeds them back to the model building module in the first step. This high-quality feedback data will be used to update the business process knowledge graph (e.g., discovering a hidden business rule), optimize the reward function or strategy in the agent simulation model (e.g., correcting the reward value of a certain behavior), and serve as new training samples to fine-tune the generative model for knowledge enhancement.

[0102] Through this complete closed loop of "planning-execution-monitoring-correction-evaluation-feedback", the entire system has the ability to continuously learn and self-evolve. Its decision-making level and model accuracy will continue to improve as the operating time increases, ultimately achieving continuous and adaptive improvement of airport ground support efficiency.

[0103] In summary, this embodiment constructs a precise, intelligent, and self-learning two-layer closed-loop optimization system through the collaborative work of four modules: model building, long-term planning, dynamic correction, decision-making, and feedback. This provides powerful, reliable, and evolvable technical support for the dynamic management and optimization of airport ground support resources.

[0104] The above are merely preferred embodiments of the present invention and are not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for dynamic optimization of airport ground support resources based on digital twins, characterized in that, The steps of this method include: The knowledge graph of airport business processes is analyzed and multimodal data is integrated to construct a generative digital twin model that reproduces and extrapolates the dynamic operation of airport ground support. Based on the generative digital twin model, multi-scenario simulation and multi-objective optimization are performed on multi-modal data to generate a long-term planning scheme covering a set future time. During the execution of the long-term planning scheme, the deviation between the actual operating data and the predicted values ​​of the scheme is monitored in real time. When the deviation is within a preset threshold, the parameters of the generative digital twin model are adaptively corrected. When the deviation exceeds the preset threshold, the long-term planning scheme is replanned to achieve dynamic adaptive adjustment of the long-term planning scheme. By combining the long-term planning scheme with real-time operational data, the system provides managers with visual decision support, evaluates the optimization effect, and uses the evaluation results as feedback to iteratively optimize the generative digital twin model.

2. The method for dynamic optimization of airport ground support resources based on digital twins according to claim 1, characterized in that, Constructing a generative digital twin model specifically involves building an airport business process knowledge graph and an intelligent agent simulation model to accurately map airport ground support scenarios. The airport business process knowledge graph is calculated using the following formula: Where G represents the business process knowledge graph; V is the set of nodes, including business entity nodes. Activity Nodes and rule nodes E is the set of edges, representing the relationships between nodes; L is the set of attribute labels for nodes and edges. Each agent in the agent simulation model The internal structure of is calculated using the following formula: in, A unique identifier for an intelligent agent; The type of intelligent agent; At the agent level; The state space of the agent; The action space of the intelligent agent; The policy function of the agent; Let be the characteristic function of the agent.

3. The method for dynamic optimization of airport ground support resources based on digital twins according to claim 2, characterized in that, The intelligent agent simulation model is defined with the following reward function, the calculation formula of which is: in, For agent i to take action a in state s to transition to a new state The reward value obtained at that time; Rewards for completing the task; Rewards for collaboration; As a reward for overall system performance; To balance the weighting coefficients of different rewards; The strategy of the agent simulation model is iteratively optimized using the Q-value update formula, specifically as follows: in, The state-action value of agent i taking action a in state s; The learning rate; Discount factor; In the new state The maximum Q value among all possible actions.

4. The method for dynamic optimization of airport ground support resources based on digital twins according to claim 3, characterized in that, The multimodal data is specifically integrated by constructing a multimodal data fusion framework to integrate multi-source heterogeneous data. This multimodal data fusion framework utilizes a self-attention mechanism to extract and fuse multimodal data features, and its calculation formula is as follows: Where Q is the query matrix; K is the key matrix; and V is the value matrix; The dimension of the key vector; Based on the reliability assessment results of each data source, the fused feature vector is dynamically weighted and updated using the following formula: in, for The fused feature vector at each time step; For the k-th data mode in The feature vector at time step; Let be the dynamic fusion weight for the k-th data modality, and .

5. The method for dynamic optimization of airport ground support resources based on digital twins according to claim 4, characterized in that, The multi-objective optimization problem defines three objectives: maximizing resource utilization, minimizing operating costs, and maximizing service quality. It is then transformed into a single-objective optimization problem through the following comprehensive optimization objective function: in, Let X be the overall objective function to be maximized; X be the resource allocation scheme. These are the weighting coefficients for each objective; The objective function is to maximize resource utilization. The objective function is to minimize operating costs; The objective function is to maximize service quality. Meanwhile, the solution process satisfies resource availability constraints, task requirement constraints, and resource transfer constraints.

6. The method for dynamic optimization of airport ground support resources based on digital twins according to claim 5, characterized in that, The adaptive correction of the parameters of the generative digital twin model specifically involves constructing a short-term dynamic correction optimization problem, which includes minimizing the deviation from long-term planning and maximizing the benefits of short-term optimization. The optimization objective of minimizing the deviation from the long-term plan is calculated using the following formula: in, This is a short-term dynamic correction scheme; This is a long-term planning scheme; and In the short-term and long-term scenarios, respectively, in terms of time... The amount of resource i allocated to task j; To optimize the time window in the short term; The formula for calculating the optimization objective of maximizing the benefits of short-term optimization is as follows: in, To improve resource utilization rate; Cost reduction rate; To improve service quality rate; These are the weighting coefficients for each benefit indicator.

7. The method for dynamic optimization of airport ground support resources based on digital twins according to claim 6, characterized in that, The specific calculation formula for evaluating the optimization effect is as follows: in, A comprehensive score for the decision-making options; to These are the weighting coefficients for each evaluation indicator; As a relative evaluation indicator for resource utilization efficiency; This is an indicator of relative changes in operating costs; For the relative change in service quality; This is an indicator of the relative change in response time. It serves as a robustness evaluation metric for the system.

8. A dynamic optimization system for airport ground support resources based on digital twins, characterized in that: include: The model building module parses the airport business process knowledge graph and integrates multimodal data to build a generative digital twin model that reproduces and extrapolates the dynamic operation of airport ground support. The long-term planning module, based on the generative digital twin model, performs multi-scenario simulation and multi-objective optimization on multimodal data to generate a long-term planning scheme covering a set future time. The dynamic correction module monitors the deviation between the actual operating data and the predicted values ​​of the plan in real time during the execution of the long-term planning scheme. When the deviation is within a preset threshold, it adaptively corrects the parameters of the generative digital twin model. When the deviation exceeds the preset threshold, a replanning of the long-term planning scheme is triggered to achieve dynamic adaptive adjustment of the long-term planning scheme. The decision-making and feedback module combines the long-term planning scheme with real-time operational data to provide managers with visual decision support, evaluate the optimization effect, and use the evaluation results as feedback to iteratively optimize the generative digital twin model.

9. The airport ground support resource dynamic optimization system based on digital twin as described in claim 8, characterized in that, The model building module, when parsing the airport business process knowledge graph, also defines a compliance penalty mechanism, the calculation formula of which is: in, For the penalty function; The model builds a knowledge-enhanced generative model in the model building module; c represents the input context. This is the embedding vector for domain knowledge; R represents the model parameters; R is the set of domain rules. This is an indicator function, which is 1 when the condition is met and 0 otherwise; Let be the penalty coefficient for rule r.

10. The airport ground support resource dynamic optimization system based on digital twin as described in claim 9, characterized in that, The dynamic correction module adaptively adjusts the weight coefficients in the optimization objective of maximizing short-term optimization benefits using the following formula: in, These are the weights of resource utilization rate, cost reduction rate, and service quality improvement rate at time t, respectively. The learning rate; Let be the overall benefit function at time t; These are the gradients of the overall benefit function with respect to each weight coefficient.

Images