Airport dynamic simulation parameter adaptive optimization method based on SAC reinforcement learning

Abstract

The invention relates to the technical field of airport analog simulation, in particular to an adaptive optimization method for airport dynamic simulation parameters based on SAC reinforcement learning, and the method comprises the steps: combining an SAC algorithm with an improved Transform model for multi-modal spatial-temporal feature extraction; static configuration and multi-time-frame real-time states of various business objects such as airport flights, airport positions, runways, guarantee vehicles and personnel are used as input, and parameters such as release intervals, sliding path priorities and resource allocation modes required by simulation are output through a strategy network; a strategy is continuously optimized in a simulation-real error closed loop by using a double-Q network and an entropy regularization mechanism, and key control parameters such as runway distribution, sliding path priority, release interval and guarantee vehicle and personnel scheduling are learned and adjusted online in real time from the simulation-real error closed loop; therefore, the simulation precision and the system robustness are ensured in a complex dynamic environment.

Description

Technical Field

[0001] This invention relates to the field of airport simulation technology, and in particular to an adaptive optimization method for airport dynamic simulation parameters based on SAC reinforcement learning. Background Technology

[0002] With the continued growth of global air transport demand, modern airports face an increasingly complex operating environment, especially in the airside, which involves multiple highly coupled dynamic processes such as flight taxiing, ground support, runway scheduling, and taxiway conflict management. These airside operations are not only affected by external uncertainties such as weather, traffic flow, and flight plan adjustments, but also depend on the spatiotemporal matching of resources such as gate allocation, runway usage, and the dispatching of support vehicles and personnel. Currently, most airport management systems rely primarily on manual experience to formulate scheduling strategies, making it difficult to quickly reconstruct and optimize scheduling plans under high-concurrency flight traffic and emergencies, resulting in significantly insufficient responsiveness.

[0003] To improve the real-time performance and simulation accuracy of airport digital twin simulation systems, it is urgent to achieve multi-dimensional, dynamically coupled adaptive optimization of simulation input parameters.

[0004] Traditional simulation methods often rely on static rules or empirical models based on historical data, which cannot effectively cope with rapid changes in operating conditions and complex constraints. The simulation prediction errors are large and lack online adjustment capabilities. Summary of the Invention

[0005] The purpose of this invention is to provide an adaptive optimization method for airport dynamic simulation parameters based on SAC reinforcement learning. This method solves the problems of traditional simulation methods that rely heavily on static rules or empirical models based on historical data, which cannot effectively cope with rapid changes in operational status and complex constraints, resulting in large simulation prediction errors and a lack of online adjustment capabilities.

[0006] To achieve the above objectives, this invention provides an adaptive optimization method for airport dynamic simulation parameters based on SAC reinforcement learning, comprising the following steps: S1: Define the input parameter types for the airport dynamic simulation model, and calibrate and constrain the discrete / continuous attributes and value ranges of each type of parameter; S2: Establish an optimal strategy model for airport business simulation parameters. By constructing a neural network that fuses multi-dimensional real-time data features, the simulation parameters composed of a time-varying strategy action space are output. S3: Establish an optimal simulation parameter strategy evaluation model, design a dual-Q network structure, and estimate the value of state-action pairs respectively; S4: Input the optimal simulation parameters output by the strategy model into the real-time simulation model, calculate the error between the multidimensional prediction index and the real data, and weight them to form the simulation accuracy error reward function; S5: Use historical data to generate a training set to pre-train the reinforcement learning model, and adaptively fine-tune and update the network based on online prediction error to achieve real-time dynamic optimization of airport airside operation simulation input parameters.

[0007] In step S1, the input parameters include: Entity object runtime data set: includes static and dynamic data of airport business objects such as flights, passengers, baggage, cargo and mail, support vehicles, support personnel, runways, aircraft stands, and weather, which are used to drive airport dynamic simulation and serve as environmental state input for reinforcement learning; Simulation time-varying strategy parameters: These are dynamically generated by the strategy model based on entity object data. They include key decision variables such as scheduling planning, operation mode, operation control, and operation object behavior. They serve as dynamic control parameters for the simulation model, adapting to different operating scenarios and correcting simulation accuracy.

[0008] The establishment of the strategy model in step S2 includes: Collect and preprocess multi-source airport business data, and integrate them into a structured feature matrix; Extract global spatiotemporal features; Construct a multi-head self-attention mechanism to fuse features across objects; The time-varying strategy parameters are output through the output layer.

[0009] Multi-source data preprocessing includes aligning and fusing multi-source airport business data, constructing structured feature matrices for different business entities, and using them as input to the strategy model.

[0010] The global spatiotemporal feature extraction includes: extracting local features of each entity object data through a one-dimensional convolutional network, and then mining the spatial-temporal features of key objects through a long short-term memory neural network to output global spatiotemporal embedding features.

[0011] Among them, the cross-object multi-head self-attention fusion mechanism includes: mapping the embedded features of each business object to a unified dimension and splicing them together; enhancing the semantic modeling between objects through an improved temporal-spatial multi-head attention mechanism; combining multi-head attention with a deep feedforward network; and constructing an improved Transformer encoder using residual connections and layer normalization.

[0012] Among them, the time-varying strategy parameter output layer predicts and outputs multiple scheduling and control strategy parameters required for airport simulation based on the features extracted by the Transformer encoder through a multi-head regression output layer; each parameter is regressed or classified through an independent network subhead to ensure that the strategy output is adapted to the multi-dimensional and heterogeneous control commands required by the airport simulation model.

[0013] In step S3, the evaluation model learns the impact of simulation parameter configuration on system operation under different states by concatenating state and action vectors into a multilayer perceptron; a dual-Q network mechanism is used to estimate the expected cumulative reward of state-action pairs, optimize the strategy and evaluate action quality, and avoid overestimation of the value function.

[0014] In step S4: The parameters output by the strategy network and the entity object data are input into the dynamic simulation model to generate simulation results that include airside operational efficiency, ground support efficiency, and passenger service efficiency. By comparing the errors between simulation results and real data, and combining preset weights, a multi-index weighted simulation accuracy error reward function is constructed to drive strategy optimization.

[0015] In step S5, a parameter correction mechanism is established based on SAC reinforcement learning: The model is pre-trained using historical data, and the historical state data is input into the policy model. Its output parameters drive the simulation model to generate a training set. Construct a dual Critic value network and optimize the Actor policy network by updating the expected value of action goals and the loss function; Based on the state-action-reward-next state sample quadruple generated by simulation, an experience replay buffer is constructed. Through closed-loop feedback, the policy network and simulation parameter model are dynamically adjusted to achieve real-time high-precision coupling between the simulation system output and the real operating environment or historical benchmark data.

[0016] This invention presents an adaptive optimization method for airport dynamic simulation parameters based on SAC reinforcement learning, constructing a digital twin system for airport airside operations with dynamic feedback and self-learning capabilities. By integrating the SAC policy network with multi-dimensional evaluation indicators from the simulation output, real-time adaptive adjustment of simulation input parameters to the actual environment and operational status is achieved. This mechanism effectively improves the consistency and accuracy between the simulation system and the real operational scenario, significantly enhances the intelligence level and real-time response capability of airport operation scheduling, and provides efficient, stable, and generalizable technical support for flight support optimization in complex scenarios, possessing significant engineering application value. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below.

[0018] Figure 1 This is a flowchart of the adaptive optimization method for airport dynamic simulation parameters based on SAC reinforcement learning, as described in this invention.

[0019] Figure 2This is a diagram of the SAC reinforcement learning model network architecture of the present invention.

[0020] Figure 3 This is the adaptive update process of dynamic simulation parameters in this invention. Detailed Implementation

[0021] The embodiments of the present invention are described in detail below. Examples of the embodiments are shown in the accompanying drawings. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, but should not be construed as limiting the present invention.

[0022] First embodiment: Please see Figures 1 to 3 ,in Figure 1 This is a flowchart of an adaptive optimization method for airport dynamic simulation parameters based on SAC reinforcement learning. Figure 2 This is the network architecture diagram of the SAC reinforcement learning model. Figure 3 It is a dynamic simulation parameter adaptive update process.

[0023] This invention provides an adaptive optimization method for airport dynamic simulation parameters based on SAC reinforcement learning, comprising the following steps: S1: Define the input parameter types for the airport dynamic simulation model, and calibrate and constrain the discrete / continuous attributes and value ranges of each type of parameter.

[0024] Specifically, in step S1, the input parameters include: Entity object runtime data set: includes static and dynamic data of airport business objects such as flights, passengers, baggage, cargo and mail, support vehicles, support personnel, runways, aircraft stands, and weather, which are used to drive airport dynamic simulation and serve as environmental state input for reinforcement learning; Simulation time-varying strategy parameters: These are dynamically generated by the strategy model based on entity object data. They include key decision variables such as scheduling planning, operation mode, operation control, and operation object behavior. They serve as dynamic control parameters for the simulation model, adapting to different operating scenarios and correcting simulation accuracy.

[0025] In this embodiment, step S1 divides the input parameters of the airport dynamic simulation model into two categories: the entity object running data set and the simulation time-varying strategy parameters.

[0026] The entity object dataset represents a collection of static and dynamic data on various airport operational objects, including flights, passengers, baggage, cargo and mail, support vehicles, support personnel, runways, aircraft stands, and weather. Static data refers to the basic information and static configurations of entity objects that do not change over time, typically obtained through engineering measurements, system documentation, or constant configurations. Dynamic data reflects the time-varying characteristics and state evolution of various operational objects in the airport operating environment, primarily achieved through real-time situational awareness and integration via various airport operational systems, monitoring equipment, and sensor networks. The entity object dataset serves two purposes: firstly, it provides timely input data for airport dynamic simulation models, driving the state evolution of the simulation model at different time steps; secondly, it serves as input to the environmental state variables of reinforcement learning policy models, providing them with the spatiotemporal characteristics of entity objects to solve for corresponding action strategies.

[0027] The time-varying strategy parameters in the simulation represent the action space variables dynamically generated by the strategy model based on the entity object dataset. These time-varying strategy parameters serve as input parameters for dynamically adjusted control strategies in the airport dynamic simulation model, adapting the airport's strategy responses to different operational scenarios and providing accuracy correction at the strategy level for the airport operation simulation model. This parameter set systematically characterizes key decision variables of different levels and types involved in airport operation simulation, covering multi-scale strategy control logic from macro-level resource scheduling to micro-level individual behavior. It mainly includes scheduling planning parameters, operation mode parameters, operation control parameters, and operational object behavior parameters. At any simulation time... The set of time-varying strategy parameters under the given conditions is:

[0028] The scheduling and planning parameters are as follows: Based on the station allocation strategy Runway allocation strategy Ground support resource scheduling strategy Personnel scheduling strategy 1. Guarantee resource scheduling strategy Passenger service scheduling strategy Composed of elements such as [list of elements], these parameters are mainly used to describe control strategy variables related to the allocation and scheduling decisions of operational resources on the airside and terminal side of the airport. These parameters address the dynamic management needs of various key resources in airport operation simulation. By setting matching rules between flights and aircraft stands, configuring runway usage strategies for arriving and departing flights, constructing scheduling models for support personnel and service vehicles, and modeling resource coordination mechanisms for service nodes such as check-in, security check, boarding, and baggage carousels, a unified scheduling expression of multiple types of resources is realized in the simulation system. This drives the airport simulation model to dynamically generate feasible, efficient, and collaborative resource allocation schemes based on operational status and strategy inputs.

[0029] The operating mode class parameters:

[0030] Configured by runway operation mode Aircraft taxiing path planning Safeguard vehicle route control Takeoff release strategy mode Entry sorting rules The composition, as the core set of strategic variables describing the airport's operation mode and flow mechanism in the simulation system, comprehensively depicts the key processes and operating rules in the airport operation process, including the runway activation direction and capacity limits, the taxiing path planning schemes for aircraft and support vehicles, the strategy classification and priority setting for flight takeoff and release, and the sorting principles for arriving flights. This parameter set systematically expresses the airport's flow mechanism and operation mode under different operating scenarios, clarifies the spatiotemporal organization relationship and scheduling logic of each operating element, and provides structured process control input for dynamic simulation.

[0031] The operation control parameters:

[0032] Control parameters of flight takeoff and landing intervals Maximum runway occupancy time limit Conflict resolution priority rules Ground support node service time And ensure the time window for resource allocation Composed of parameters such as time control and resource allocation, this system represents a set of control strategy variables that characterize the constraints of various transportation entities and support resources in time control and conflict management during airport operations. It systematically describes the time control, resource occupation boundaries, and conflict management rules in airport operations, and models and configures operational control rules.

[0033] The behavior class parameters of the running object:

[0034] Covering aircraft , and support vehicles Passengers ,luggage Goods and mail and support personnel This includes a set of strategy variables for various airport business objects, used to describe the micro-behavioral characteristics of multiple agents in airport operation simulation. It constructs expected speed, maximum and acceptable acceleration / deceleration, turning radius, path selection strategy, task service duration, waiting time distribution, and behavioral preference rules for each state stage. These are used to define the dynamic control boundaries and decision-making basis for individuals in spatial movement, task execution, resource utilization, and behavioral response processes, supporting the behavioral modeling requirements of operating entities in different time periods, spatial regions, and state transitions.

[0035] S2: Establish an optimal strategy model for airport business simulation parameters. By constructing a neural network that fuses multi-dimensional real-time data features, the simulation parameters are output, which consist of a time-varying strategy action space.

[0036] S21: Collect and preprocess multi-source airport business data, and integrate them into a structured feature matrix.

[0037] In this implementation, multi-source operational data generated during airport operations, including flight schedule data, aircraft positioning data, support resource scheduling information, and air traffic control meteorological data, are collected and preprocessed. These data are aligned and fused according to a unified time base and spatial identifier to construct a structured feature information matrix for different operational entities. This matrix serves as input to a reinforcement learning strategy model, enabling the learning and optimization of decision-making strategies under complex airport operational conditions. Each entity type is treated as a data dimension, fusing its static attributes and dynamic behavioral characteristics during operation to form a multimodal feature representation with time-series characteristics.

[0038] in, This is a feature information matrix for the dataset of all entity objects, including various airport business objects such as flights, passengers, baggage, cargo and mail, support vehicles, support personnel, runways, aircraft stands, and weather. Each type of object input contains a multimodal operational data set including resource configuration, trajectory location, operational status, and support plans. The dataset also consists of... It consists of a sequence of historical time frames, with each frame being a multidimensional vector. The input tensor for a single object is:

[0039] in, express The object at time The static configuration, operational status, scheduling information, and other multimodal features of the flight data include aircraft number, airline, aircraft type, assigned gate, assigned runway, real-time flight path, airspace route, surface taxiing route, flight node status, departure time, and arrival time. These features can comprehensively characterize the status and behavioral logic of flights within the airport operation system.

[0040] S22: Extract global spatiotemporal features.

[0041] In this embodiment, a local temporal convolutional layer is established, and for each type of airport business object, the input data tensor is... One-dimensional convolutional network using shared parameters Local feature extraction is performed on the spatial dimension features within each time frame to capture the local correlation and structural patterns among the multimodal features contained in the object at a single moment, and the convolutional embedding features of each object dataset are also analyzed. Represented as: , in, right Perform one-dimensional convolution along the time dimension. This is a non-linear activation function. Based on this, to further model the temporal dependencies and dynamic evolution of objects during operation, a long short-term memory neural network is introduced. Global temporal modeling is performed on the convolutional feature sequences. Through in-depth spatial-temporal feature mining of multiple key objects such as flights, baggage, support vehicles, and weather, a global spatiotemporal embedding feature tensor is output. ,

[0042] S23: Construct a multi-head self-attention mechanism to fuse features across objects.

[0043] In this implementation, a multi-head self-attention fusion mechanism across objects is constructed to uniformly map the embedded representations of various business objects to a unified dimension. , recorded as After being concatenated, the data is sent to the multi-object fusion Transformer module:

[0044] Utilizing an improved temporal-spatial multi-head attention mechanism:

[0045] in, Indicates timing position offset. Learnable embedding biases representing object categories enhance semantic modeling capabilities between objects. Simultaneously, an improved Transformer encoder is built: combining multi-head attention with a deep feedforward network (FFN), employing residual connections and layer normalization.

[0046] in, This is a multi-head self-attention mechanism. It is a feedforward neural network used to independently perform feature transformation and nonlinear enhancement on the representation of the object at each time step. Perform normalization operation, the first The input feature tensor of the layer is transformed into the first layer. The output feature tensor of the layer.

[0047] S24: Output time-varying strategy parameters through the output layer.

[0048] In this embodiment, in the policy parameter output layer, the model feeds the temporal feature representation extracted by the Transformer encoder into the multi-head regression output layer to predict and output multiple scheduling and control policy parameters required for airport simulation. These policy parameters constitute the action space of the reinforcement learning model, and the final output tensor is activated by an activation function. The last frame or average of all time frames is used as the policy prediction output, and regression predictions are performed on different types of policy variables respectively:

[0049] in, This represents the pooling operation on the hidden states of the time series, forming the final constructed state vector which is then input into the policy network. For a multilayer perceptron policy network, the state vector is mapped to the parameters of the action distribution. It is the mean of the action of the i-th parameter. The standard deviation is denoted as . Let be the activation function for the action space. To introduce randomness into the strategy, The i-th output represents the policy action value, which serves as the simulation input parameter. The output types include classification probability, multi-class selection, and continuous value estimation. Each parameter is regressed or classified through an independent network subheading to ensure that the policy output adapts to the multi-dimensional and heterogeneous control commands required by the airport simulation model.

[0050] S3: Establish an optimal simulation parameter strategy evaluation model, design a dual-Q network structure, and estimate the value of state-action pairs respectively.

[0051] In this embodiment, step S3 specifically establishes an optimal simulation parameter strategy evaluation model, namely a Critic network in reinforcement learning. By concatenating state and action vectors and inputting them into a multilayer perceptron, the network can learn the impact of various simulation parameter configuration schemes on the system's performance under different business states. To avoid overestimation of the value function, a dual Q-network mechanism is introduced to construct two action value functions, which estimate the current state respectively. Next action The expected cumulative return that can be obtained is used to optimize the strategy and evaluate the quality of the actions:

[0052] Among them, state Characterizing the current airport operating environment (such as runway occupancy, taxi queues, weather conditions, etc.), actions Characterize the simulation parameter configuration scheme (such as sliding interval, priority sorting, release window, etc.). State vector and action vectors The splicing of these elements forms the input to the evaluation network. and These are two independent Critic networks, typically structured as multilayer perceptrons, with an output of a scalar. .

[0053] S4: Input the optimal simulation parameters output by the strategy model into the real-time simulation model, calculate the error between the multidimensional prediction index and the real data, and weight them to form the simulation accuracy error reward function.

[0054] In this embodiment, step S4 inputs the optimal simulation parameter strategy action into the real-time simulation model, compares the simulation output running index with the real data of the next time step, and constructs a reward function for the simulation accuracy error.

[0055] The system operates within each simulation cycle (the cycle is...). Real-time collection of current environmental status Various simulation parameters. These will be available at time [time]. Simulation parameter action set output by the policy network The airport entity object dataset is input into the airport dynamic simulation model to obtain a set of simulation result index vectors. , and the simulation system at time Real data output metrics Compare the data and collect changes in operational performance indicators. The reward function is as follows:

[0056] Each component These are simulation prediction performance evaluation indicators. For example, airside operational efficiency measures the smoothness of airport airspace and taxiway operations and conflict control capabilities, including indicators such as runway utilization balance, aircraft ground taxiing time, and the number of taxiing conflicts. Ground support efficiency assesses the timeliness of flight support processes and the rationality of resource allocation, including indicators such as on-time rate of support vehicles, support resource occupancy rate, and flight delay time. Passenger service efficiency reflects the level of passenger support in various service processes within the terminal, including indicator dimensions such as the number of people queuing and queuing time distribution at service facilities, arrival and departure passenger distribution, and the time spent on each passenger process. By introducing prediction error terms for different types of indicators, these methods can be used to further enhance the effectiveness of simulation prediction. With positive improvement feedback items The cumulative weighted sum is the reward function. , These are the weighting factors for each indicator.

[0057] S5: Use historical data to generate a training set to pre-train the reinforcement learning model, and adaptively fine-tune and update the network based on online prediction error to achieve real-time dynamic optimization of airport airside operation simulation input parameters.

[0058] In this embodiment, step S5 establishes a parameter correction mechanism based on the SAC reinforcement learning network to dynamically correct the simulation input parameters output by the policy network.

[0059] A large training set is generated using historical data for pre-training of reinforcement learning models. Historical airport real-time status data is fed into the airport business simulation parameter strategy model. The output simulation input parameter results are used to provide simulation inference and prediction to the airport business simulation model, generating a large training set.

[0060] Update the Critic (value network) in SAC and construct two Critic networks. and The state-action value function is estimated separately. To enhance the exploratory nature and robustness of policy generation, a maximum entropy reinforcement learning framework is adopted, which introduces an entropy regularization term to update the expected action value at the current time step. :

[0061] In the formula, The state at the next moment; A set of action spaces sampled from the action distribution of the airport's operational status at the next moment; Use the smaller of the two target values ​​(Critic) to prevent overestimation; This represents the conditional distribution of the output parameter actions; θ represents the entropy temperature coefficient, which adjusts the trade-off between the exploratory nature of the strategy and the maximization of value; θ represents all parameters of the Actor network. Discount factor; This is the current instant reward.

[0062] Furthermore, by minimizing the mean squared error for each of the two Critic networks, their loss functions are calculated as follows:

[0063] In the formula, This represents the loss function of the Critic network, which is updated using gradient descent. ; Furthermore, the objective function for updating the Actor policy network in SAC is:

[0064] In the formula, Let represent the policy objective function that maximizes the expected Q-value after entropy regularization; Indicates the state From the experience replay pool The expected value of sampling in the middle; Indicates the state Next, strategy Output action Expectations; This indicates the output of the policy network in the state. Next action The probability density value is sampled and the gradient is calculated to update the Actor parameters. Finally, a closed-loop feedback mechanism of "simulation-evaluation-update-resimulation" is constructed. In this mechanism, each generated set of simulation state, action, reward, and next state sample quadruple is generated. All data is stored in real-time in the replay buffer, serving as the foundation for subsequent updates to the policy network and value network. By alternately executing update operations for the Critic evaluation network and the Actor policy network, the system continuously compares the errors between the simulation output results and the actual operating conditions or historical calibration data. This optimizes the calibration action policy and state value estimation, thereby achieving gradual convergence and dynamic optimality of the policy. The simulation input parameters can be dynamically adjusted according to the environment and operating conditions, ensuring that the dynamic simulation model remains highly consistent with the real environment. This highly coupled feedback process significantly improves the adaptability and prediction accuracy of the simulation modeling.

[0065] The above-disclosed embodiments are merely one or more preferred embodiments of this application and should not be construed as limiting the scope of this application. Those skilled in the art can understand that all or part of the processes for implementing the above embodiments and equivalent changes made in accordance with the claims of this application still fall within the scope of this application.

Claims

1. An adaptive optimization method for airport dynamic simulation parameters based on SAC reinforcement learning, characterized in that, Includes the following steps: S1: Define the input parameter types for the airport dynamic simulation model, and calibrate and constrain the discrete / continuous attributes and value ranges of each type of parameter; S2: Establish an optimal strategy model for airport business simulation parameters. By constructing a neural network that fuses multi-dimensional real-time data features, the simulation parameters composed of a time-varying strategy action space are output. S3: Establish an optimal simulation parameter strategy evaluation model, design a dual-Q network structure, and estimate the value of state-action pairs respectively; S4: Input the optimal simulation parameters output by the strategy model into the real-time simulation model, calculate the error between the multidimensional prediction index and the real data, and weight them to form the simulation accuracy error reward function; S5: Use historical data to generate a training set to pre-train the reinforcement learning model, and adaptively fine-tune and update the network based on online prediction error to achieve real-time dynamic optimization of airport airside operation simulation input parameters.

2. The adaptive optimization method for airport dynamic simulation parameters based on SAC reinforcement learning as described in claim 1, characterized in that, In step S1, the input parameters include: Entity object runtime data set: includes static and dynamic data of airport business objects such as flights, passengers, baggage, cargo and mail, support vehicles, support personnel, runways, aircraft stands, and weather, which are used to drive airport dynamic simulation and serve as environmental state input for reinforcement learning; Simulation time-varying strategy parameters: These are dynamically generated by the strategy model based on entity object data. They include key decision variables such as scheduling planning, operation mode, operation control, and operation object behavior. They serve as dynamic control parameters for the simulation model, adapting to different operating scenarios and correcting simulation accuracy.

3. The adaptive optimization method for airport dynamic simulation parameters based on SAC reinforcement learning as described in claim 1, characterized in that, The establishment of the policy model in step S2 includes: Collect and preprocess multi-source airport business data, and integrate them into a structured feature matrix; Extract global spatiotemporal features; Construct a multi-head self-attention mechanism to fuse features across objects; The time-varying strategy parameters are output through the output layer.

4. The adaptive optimization method for airport dynamic simulation parameters based on SAC reinforcement learning as described in claim 3, characterized in that, Multi-source data preprocessing includes aligning and fusing multi-source airport business data, constructing structured feature matrices for different business entities, and using them as input to the strategy model.

5. The adaptive optimization method for airport dynamic simulation parameters based on SAC reinforcement learning as described in claim 3, characterized in that, Global spatiotemporal feature extraction includes: extracting local features of each entity object data through a one-dimensional convolutional network, and then mining the spatial-temporal features of key objects through a long short-term memory neural network to output global spatiotemporal embedding features.

6. The adaptive optimization method for airport dynamic simulation parameters based on SAC reinforcement learning as described in claim 3, characterized in that, The cross-object multi-head self-attention fusion mechanism includes: mapping the embedded features of each business object to a unified dimension and concatenating them; enhancing the semantic modeling between objects through an improved temporal-spatial multi-head attention mechanism; combining multi-head attention with a deep feedforward network; and constructing an improved Transformer encoder using residual connections and layer normalization.

7. The adaptive optimization method for airport dynamic simulation parameters based on SAC reinforcement learning as described in claim 3, characterized in that, The time-varying strategy parameter output layer predicts and outputs multiple scheduling and control strategy parameters required for airport simulation based on features extracted by the Transformer encoder through a multi-head regression output layer. Each parameter is regressed or classified through an independent network subhead, ensuring that the strategy output adapts to the multi-dimensional and heterogeneous control commands required by the airport simulation model.

8. The adaptive optimization method for airport dynamic simulation parameters based on SAC reinforcement learning as described in claim 1, characterized in that, The evaluation model in step S3 learns the impact of simulation parameter configuration on system operation under different states by concatenating state and action vectors and inputting them into a multilayer perceptron. A dual-Q network mechanism is used to estimate the expected cumulative reward of state-action pairs, optimize the strategy, and evaluate action quality, thus avoiding overestimation of the value function.

9. The adaptive optimization method for airport dynamic simulation parameters based on SAC reinforcement learning as described in claim 1, characterized in that, In step S4: The parameters output by the strategy network and the entity object data are input into the dynamic simulation model to generate simulation results that include airside operational efficiency, ground support efficiency, and passenger service efficiency. By comparing the errors between simulation results and real data, and combining preset weights, a multi-index weighted simulation accuracy error reward function is constructed to drive strategy optimization.

10. The adaptive optimization method for airport dynamic simulation parameters based on SAC reinforcement learning as described in claim 1, characterized in that, In step S5, a parameter correction mechanism is established based on SAC reinforcement learning: The model is pre-trained using historical data, and the historical state data is input into the policy model. Its output parameters drive the simulation model to generate a training set. Construct a dual Critic value network and optimize the Actor policy network by updating the expected value of action goals and the loss function; Based on the state-action-reward-next state sample quadruple generated by simulation, an experience replay buffer is constructed. Through closed-loop feedback, the policy network and simulation parameter model are dynamically adjusted to achieve real-time high-precision coupling between the simulation system output and the real operating environment or historical benchmark data.

Images