Satellite ground station antenna intelligent scheduling method in multi-satellite concurrent scenario

Abstract

The application discloses a kind of satellite ground station antenna intelligent scheduling methods under multi-star concurrent scene, including, construct digital twinborn environment, physical and meteorological modeling is carried out to antenna electromechanical servo dynamic response and communication link quality;High-dimensional state information is converted into dynamic graph structure, offline training is carried out in twinborn environment using deep reinforcement learning model, through multi-objective composite reward and course learning mechanism, learn global scheduling strategy giving consideration to throughput, priority and mechanical wear;When online application, carry out short-term domain pre-inference in combination with real perception state, output scheduling instruction after interception by hard constraint safety verification, and use real interactive data to carry out incremental fine-tuning to model.The application effectively overcomes the dimension disaster and idealization assumption of traditional algorithm, realizes the antenna automatic optimal scheduling of high reliability, anti-collision and self-adapting equipment aging in dynamic scene.

Description

Technical Field

[0001] This invention relates to the field of satellite communication and ground station automation control technology, and in particular to a method for intelligent scheduling of satellite ground station antennas in multi-satellite concurrent scenarios. Background Technology

[0002] With the rapid development of low-Earth orbit satellite constellation technology, the construction of global satellite internet has entered a period of explosive growth, and satellite ground stations are facing unprecedented challenges of high frequency and multi-task concurrency. In multi-satellite concurrent scenarios, how to scientifically allocate limited ground station antenna resources to achieve optimal communication link coverage and maximize data transmission efficiency has become a major research direction in the field of satellite ground system control. Existing ground station scheduling schemes are undergoing a transformation from manual experience-based planning to automated search algorithms. Early technologies were mainly based on prior rules or simple time-series allocation, and subsequently evolved into traditional heuristic search methods such as genetic algorithms and ant colony algorithms. These methods are feasible to a certain extent in scenarios with small-scale tasks and relatively static environmental parameters, providing a fundamental guarantee for improving the automation level of ground station operations.

[0003] However, existing technologies still exhibit significant limitations when dealing with the dynamic concurrent tasks of ultra-large-scale satellite constellations. First, traditional modeling methods are often based on idealized geometric link models, easily overlooking atmospheric precipitation attenuation, electromagnetic environment fluctuations, and the non-instantaneous dynamic response of antenna electromechanical servo systems during high-speed pointing switching. This leads to execution deviations or excessively rapid mechanical wear in the actual physical environment. Second, when handling complex topological dependencies between multiple satellites and antennas, existing algorithms often fall into the curse of dimensionality due to a lack of deep representation of entity characteristics and their potential associations, making it difficult to output the globally optimal solution within a short decision-making cycle. Furthermore, existing offline scheduling models typically use fixed evaluation metrics, lacking the ability to balance and optimize multiple objectives such as task priority, load balancing, and data freshness. Moreover, the absence of online security verification mechanisms and incremental learning features results in insufficient robustness in the face of sudden tasks or abnormal equipment conditions, making it difficult to meet the practical requirements of intelligent and highly reliable operation of modern satellite ground stations. Summary of the Invention

[0004] The purpose of this section is to outline some aspects of embodiments of the present invention and to briefly describe some preferred embodiments. Simplifications or omissions may be made in this section, as well as in the abstract and title of this application, to avoid obscuring the purpose of these documents; however, such simplifications or omissions should not be construed as limiting the scope of the invention.

[0005] In view of the aforementioned existing problems, this invention is proposed. Therefore, this invention provides an intelligent scheduling method for satellite ground station antennas in multi-satellite concurrent scenarios to solve the problems mentioned in the background art.

[0006] To address the aforementioned technical problems, this invention provides the following technical solution: a method for intelligent scheduling of satellite ground station antennas in a multi-satellite concurrent scenario, comprising:

[0007] S1: Construct a digital twin scheduling environment, model the orbital dynamics of ground station antennas, multiple satellites, and the quality of communication links between antennas and satellites through the environment, and predict the dynamic changes of the quality of the communication links in the future time period;

[0008] S2: Based on a deep reinforcement learning model, it is trained offline in the digital twin scheduling environment. Through simulated interaction with the environment, it learns a scheduling strategy that maximizes a multi-objective composite reward under preset constraints.

[0009] S3: Using the trained deep reinforcement learning model, generate and output antenna scheduling instructions for the current decision cycle online based on the real-time perceived status of ground stations, satellites, and missions.

[0010] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0011] 1. This invention transforms the concurrent scheduling problem of multiple satellites and multiple antennas into a dynamic graph structure. By combining graph attention networks with deep reinforcement learning, the system can adaptively handle the on-demand access or removal of nodes (satellites), thereby understanding spatial topological dependencies.

[0012] 2. Existing scheduling algorithms are often based on the ideal assumption of instantaneous switching. This invention constructs a digital twin environment, integrating the non-instantaneous dynamic response of the antenna servo system (angular velocity / angular acceleration constraints) and link attenuation prediction caused by weather / spectrum. By converting the mechanical switching time into a switching penalty in multi-objective rewards, the model is forced to learn the logic of trading space for time, avoiding equipment vibration and wear caused by issuing commands that exceed physical limits.

[0013] 3. In addition, to address the uncertainty risks that may arise from black-box exploration of reinforcement learning models, this invention filters short-sighted suboptimal solutions through short-time domain pre-deduction when generating instructions online, and performs hard boundary interception through a model-independent safety sandbox. This enables the decoupling of the model's efficient optimization capability from the safety of industrial control, ensuring that even under extreme conditions, the antenna will not exceed physical limits, thereby meeting the stringent requirements for high reliability in complex fields such as aerospace telemetry and control. Attached Figure Description

[0014] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Wherein:

[0015] Figure 1 This is a flowchart illustrating the overall process of intelligent scheduling method for satellite ground station antennas in a multi-satellite concurrent scenario, as described in one embodiment of the present invention. Detailed Implementation

[0016] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0017] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0018] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0019] Furthermore, in the description of this invention, it should be noted that the terms "upper," "lower," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. These terms are used solely for the convenience of describing the invention and for simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the invention. In addition, the terms "first," "second," or "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0020] Example 1

[0021] Reference Figure 1 This is the first embodiment of the present invention, which provides a method for intelligent scheduling of satellite ground station antennas in a multi-satellite concurrent scenario, including:

[0022] S1: Construct a digital twin scheduling environment to model the orbital dynamics of ground station antennas, multiple satellites, and the quality of communication links between antennas and satellites, and predict the dynamic changes in communication link quality over future time periods.

[0023] Furthermore, the digital twin environment, by accessing the two-line orbital elements (TLE) or precise orbital data of the satellites, utilizes high-precision orbital extrapolation algorithms (such as SGP4 or HPOP) to reconstruct the motion trajectories of multiple satellites relative to the ground station in a virtual three-dimensional space. During this modeling process, the digital twin environment calculates the azimuth, elevation, and instantaneous distance of the satellites in the local coordinate system of the ground station in real time, which is used to define the start time of each satellite entering the visible window above the obstruction angle of the ground station.

[0024] It should be noted that through the above modeling process, the scheduling system can identify the overlapping areas of multiple satellite windows on the time axis, thereby transforming the scheduling requirements of multiple satellites concurrently into a deterministic time interval conflict problem, providing a basic search boundary for resource allocation.

[0025] Furthermore, in order to resolve the contradiction between the idealized assumption of instruction-to-execution in traditional scheduling schemes and the rotation delay of physical equipment, this invention simulates the servo dynamic characteristics of the antenna mechanical structure through digital means.

[0026] Specifically, the modeled digital twin environment integrates the rotational inertia of the antenna base and its moving parts, the maximum angular velocity limit of the drive motor, and the angular acceleration curve constraints during the start-up and shutdown phases. When the scheduling strategy requires the antenna to turn from the currently observed satellite to the next satellite, the digital twin environment automatically calculates the physical switching time and trajectory required for the antenna to complete the elevation and azimuth rotation based on the target pointing span.

[0027] It should be noted that the advantage of doing this is that it quantifies the mechanical fatigue risk of the antenna and the time cost of pointing switching into soft and hard constraints for scheduling. This modeling approach effectively avoids the scheduling algorithm issuing physically impossible rapid switching commands, thereby reducing frequent sudden stops and oscillations in the servo system, extending equipment lifespan, and improving tracking accuracy.

[0028] Furthermore, this invention constructs a high-fidelity channel quality assessment model in a digital twin environment to predict the real-time signal-to-noise ratio (SNR) or link margin of the communication link between the antenna and the satellite. This assessment model not only considers free-space loss due to variations in satellite distance but also integrates the attenuation effect of the environment on radio waves, specifically in the following two ways:

[0029] The first method involves accessing real-time meteorological forecast data such as rainfall rate and cloud thickness in the area where the ground station is located, and combining this with the International Telecommunication Union (ITU)’s empirical model for predicting radio wave propagation (e.g., the ITU-R P.618 model). The rainfall rate parameter is then substituted into a specific polarization attenuation formula for a specific frequency (e.g., the Ka / Ku band) to quantify the additional energy loss (i.e., signal attenuation decibels) caused by rainfall to a specific spatial pointing link.

[0030] The second approach utilizes historical spectrum scanning data and surrounding electromagnetic environment records to construct a prediction model based on Long Short-Term Memory (LSTM) networks. This model predicts whether there is a risk of co-channel interference from ground base stations or other radio frequency equipment in a specific location and time period, and converts this risk into an interference probability value between 0 and 1.

[0031] In one feasible process of this embodiment, a prediction model based on a Long Short-Term Memory (LSTM) network is adopted. The historical spectrum scan sequence matrix (including frequency point, time, and Received Signal Strength Indication RSSI) within a set time period is used as the time series input. After the time series feature extraction of the LSTM layer, the probability value of a specific frequency band being subject to co-channel interference in the future visible window is output by the fully connected layer. This probability value is converted into a penalty factor and written into the edge attribute of the dynamic graph structure.

[0032] It should be noted that the above two influence methods enable scheduling instructions to have a priori implications. For example, if it is predicted that a high-priority satellite will experience link attenuation due to heavy rainfall during the upcoming window period, causing the link to become unavailable, the system can allocate antenna resources in advance to another satellite with better link quality, thereby achieving an optimal balance in data transmission efficiency.

[0033] Furthermore, the digital twin environment extracts and outputs all the above modeling parameters as a multi-dimensional state dataset containing dynamic parameters of the antenna entity and the satellite entity, which serves as the basic data for constructing the dynamic graph structure.

[0034] It should be explained that this multidimensional state dataset includes the current mechanical state of the antenna (e.g., current pointing, motor load), the satellite's motion trend (e.g., remaining time until the end of the window), and the predicted evolution trend of link quality. This multidimensional state dataset will serve as input to the deep reinforcement learning model, establishing a strong coupling between the physical environment and the intelligent algorithm. This ensures that the deep reinforcement learning model learns not blind mathematical rules, but rather understands the physical logic of trading space for time and link stability for task continuity in ground station scheduling, thereby making the online generated scheduling instructions practical and reliable.

[0035] S2: Based on a deep reinforcement learning model, it is trained offline in a digital twin scheduling environment. Through simulated interaction with the environment, it learns a scheduling strategy that maximizes a multi-objective composite reward under preset constraints.

[0036] It should be noted that, in order to enable the deep reinforcement learning model to understand the complex spatial topology of multi-star concurrency and make forward-looking decisions, this invention completes the construction and offline training of the model through dynamic graph structured representation, multi-objective composite reward design and course learning mechanism.

[0037] Furthermore, in multi-satellite concurrent scenarios, the number, position, and visibility of satellites change dynamically over time, and traditional fixed-dimensional state vectors may not be able to effectively represent such abrupt changes in the environment. Therefore, this invention constructs the high-dimensional state information output from the digital twin environment in S1 into a dynamic graph structure containing two types of nodes in real time.

[0038] Specifically, in this dynamic graph structure, nodes are divided into antenna nodes and satellite-mission nodes. The antenna node's characteristic attributes include its current azimuth, elevation, mechanical load status, and whether it is in idle / standby mode. The satellite-mission node's characteristic attributes include the satellite's predicted three-dimensional coordinates, the priority of its current mission, the total amount of data to be transmitted, and the remaining visible time within the antenna's viewing window. A connection edge is generated between two satellites only when a satellite enters the visible space of an antenna. The characteristic attributes of this edge characterize the potential communication relationship between the two, specifically including the real-time communication link quality predicted in S1 (e.g., signal-to-noise ratio affected by weather), and the non-instantaneous mechanical rotation time required for the antenna to switch to that satellite.

[0039] It should be noted that by constructing a dynamic graph, the system can transform the spatial geometric and physical constraints of which antenna can see which satellite, the communication quality, and the switching cost in the physical world into a topological connection network that can be directly read by a graph neural network. This solves the problem of the model input dimension not being fixed due to the real-time change in the number of concurrent satellites.

[0040] Furthermore, the deep reinforcement learning model employs a neural network architecture that combines a graph neural network module and a policy output module, as follows:

[0041] Graph Feature Extraction Layer (Information Aggregation): The model first uses a multi-layer graph attention network to process the dynamic graph structure constructed above. Through the message passing mechanism in the network, each node can automatically collect and fuse the state information of its connected neighboring nodes. For example, after feature aggregation, an antenna can perceive the congestion level and link competition status of all candidate satellites within its coverage area, thereby outputting a node feature vector containing global spatial dependencies.

[0042] Policy Output Layer (Action Decision): After the model completes graph structure processing, for each idle antenna node, the policy output module concatenates and fuses the antenna's feature vector with the feature vectors of all candidate satellite-task nodes within its visible range. These fused feature data are then fed into a multilayer perceptron network (MLP) for nonlinear mapping and, through a probability normalization function (e.g., Softmax), outputs an action selection probability distribution for that antenna node. It's important to note that this probability distribution represents the recommendation index for the current antenna to select each visible satellite as its service target (the sum of the probabilities of all candidate satellites is 1). Finally, the system determines the specific scheduling action for each idle antenna based on this probability distribution, such as directly selecting the target satellite with the highest probability, or exploring new scheduling schemes through probability sampling during the training phase.

[0043] In addition, this invention defines a discrete action space for the dynamic graph structure. For each idle antenna i in the dynamic graph structure, its action can be represented as follows: Where 0 indicates standby, and 1 to N indicate selecting the Nth visible satellite in the graph structure for tracking. To avoid the conflict logic problem of multiple antennas simultaneously scheduling the same single-beam satellite, the policy output layer introduces an action masking mechanism before the probability normalization function. This mechanism forces the selection probability of satellite nodes already selected by other antennas to negative infinity (i.e., output probability is 0), thereby masking invalid concurrent conflict actions at the algorithm's underlying level.

[0044] Furthermore, since the optimization direction of deep reinforcement learning is entirely determined by the reward function, this invention designs a multi-objective composite reward scoring system to prevent the model from blindly pursuing data throughput and causing excessive equipment wear or missing urgent tasks. After each simulated interaction between the deep reinforcement learning model and the digital twin environment, the environment feeds back a comprehensive score to the model based on the scheduling results. This score integrates the following four dimensions:

[0045] Data throughput reward: This score is a positive reward, and its value is positively correlated with the total amount of data successfully transmitted by the system during the current decision-making cycle. It aims to guide the model to maximize the utilization of space-ground communication resources.

[0046] Event Priority Reward: This score is a positive step reward. A high score is given when the model successfully schedules and completes a high-priority task (e.g., issuing an emergency telemetry command), ensuring that critical tasks are not crowded out by a large number of low-value routine tasks.

[0047] Load balancing reward: This reward is calculated based on the difference in historical utilization of each antenna (such as the inverse of the utilization variance). Its purpose is to encourage the scheduling strategy to distribute tasks evenly to each antenna in the ground station network, preventing a single antenna from overheating and crashing due to continuous high-intensity operation.

[0048] Switching Penalty: This is a negative deduction item. The non-instantaneous completion time and trajectory of the antenna calculated in S1 are converted into mechanical loss costs. It should be clear that large-scale rapid antenna turns or frequent switching will result in severe deductions, thus forcing the model to learn the logic of trading space for time, prioritizing scheduling sequences with smooth trajectories and short mechanical time consumption.

[0049] It is worth noting that multi-objective composite rewards The calculation is performed using a weighted scalarization formula, as detailed below:

[0050] in, The preset hyperparameter weights (in this embodiment, set to) It can also be set according to the actual situation, but its sum shall not exceed 1). The data throughput reward is calculated as a normalized value of the number of MB successfully transmitted in the current period. As an event priority reward, when a high-priority task is completed, Otherwise ; The load balancing reward is a negative number representing the variance of all antenna utilization. The penalty for switching is calculated in proportion to the non-instantaneous mechanical rotation time; the longer the switching time, the more points are deducted.

[0051] In one possible implementation of this embodiment, the deep reinforcement learning model uses a proximal policy optimization (PPO) algorithm based on an Actor-Critic architecture for parameter updates.

[0052] Specifically, the aforementioned graph feature extraction layer and policy output layer are combined to form an Actor network, which outputs the probability distribution of antenna scheduling actions. Simultaneously, a Critic network with a similar network structure but outputting a single scalar value is constructed in parallel to evaluate the expected total reward value of the current graph structure state. In each simulated interaction with the digital twin environment, the comprehensive score (reward) of the environment feedback is used to calculate the advantage function. The model uses the Adam optimizer, based on this advantage function and employing a pruning objective function that limits the policy update magnitude, to perform gradient ascent updates on the parameters of the Actor network. The mean squared error loss function is then used to perform gradient descent updates on the parameters of the Critic network. During offline training, a base learning rate (typically 1×10⁻⁶) is set. -4 The batch size is dynamically adapted based on the number of concurrent satellites, and the temporal correlation is disrupted through an experience replay mechanism until the neural network parameters converge to the preset loss threshold range.

[0053] It should also be noted that during the parameter update of the PPO algorithm, the Actor network uses this pruning objective function to limit the ratio of the new and old strategies. Within the range ( Typically, a value of 0.2 is used to prevent excessively large policy update steps from causing training crashes. This advantage function is calculated by the Critic network using generalized advantage estimation (GAE). Furthermore, to encourage the model to explore in the digital twin environment, an information entropy reward from the policy distribution can be incorporated into the loss of the Actor network.

[0054] It should be noted that by adopting the PPO algorithm and the corresponding network update logic, the training stability of the model when exploring the optimal scheduling strategy is ensured, the algorithm crash caused by excessively large strategy update steps is avoided, and the feasibility of the antenna intelligent scheduling scheme of this invention is ensured.

[0055] Furthermore, if an untrained model is directly allowed to explore in a twin environment containing complex rainfall attenuation, stringent mechanical constraints, and a massive number of satellites, the model is highly likely to fail to converge due to the excessive complexity of the environment. Therefore, this embodiment employs a course-based learning mechanism for offline training, simulating the human learning and cognitive process from easy to difficult. Specifically, it is divided into the following three stages:

[0056] Basic Phase (Simple Scenario): First, an idealized, simple scenario is set up in the digital twin environment, characterized by a small number of concurrent satellites, no weather attenuation, no radio frequency interference, and negligible antenna switching latency. In this phase, the model quickly learns the basic timing allocation capability of prioritizing connections within line of sight and ensuring connections are made to high-priority targets first.

[0057] Advanced Stage (Medium Scenario): Gradually increase the concurrent satellite density in the environment, introducing resource competition conflicts caused by the high overlap of multiple satellite view windows. In this stage, the model learns to resolve competition conflicts between multiple antennas through trial and error, utilizing probability distributions to form a locally optimal allocation strategy.

[0058] Mature Stage (Complex Scenarios): Fully activate the electromechanical servo dynamic response constraints and the attenuation effect of meteorological / spectral interference prediction in S1. Since the model already has basic scheduling capabilities, it can now focus on optimizing advanced strategies, such as proactively switching antennas in advance when facing expected heavy rainfall, and planning globally near-optimal scheduling actions while meeting mechanical safety red lines.

[0059] It should be noted that after tens of thousands of iterations of state perception-action output-reward scoring in a digital twin environment, the network parameters in the deep reinforcement learning model will eventually converge and solidify into an intelligent scheduling strategy capable of coping with extreme multi-star concurrency and complex physical environment disturbances.

[0060] S3: Utilizing a trained deep reinforcement learning model, antenna scheduling instructions are generated and output online based on real-time sensing of ground station, satellite, and mission status for the current decision cycle.

[0061] It should be noted that while offline training endows deep reinforcement learning models with global planning capabilities in virtual environments, real physical environments inevitably contain some sudden disturbances (such as emergency observation tasks inserted temporarily, or increased resistance of servo motors due to sudden gusts of wind). Therefore, in order to bridge the gap between digital simulation and physical reality and ensure the absolute safety and high reliability of online scheduling, this invention addresses the aforementioned problem of sudden disturbances through dynamic state reconstruction, short-time domain pre-simulation, boundary security interception, and online policy fine-tuning.

[0062] Furthermore, during each physical decision-making cycle (e.g., every 10 seconds), the system collects real-time physical conditions through the ground station's sensor network, telemetry and control baseband, and real-time ephemeris data interface. This data includes the actual mechanical pointing angles of each antenna, the current actual signal-to-noise ratio, and the latest satellite mission list. The system synchronously inputs this real-time data into the digital twin environment constructed by S1 and, following the aforementioned method, reconstructs the dynamic graph structure of the current moment online, using this as input for real-time inference by the deep reinforcement learning model.

[0063] It should be noted that by updating the dynamic graph structure, even if a satellite suddenly goes offline due to a malfunction, or a ground station antenna is manually locked due to temporary maintenance (i.e., these nodes and edges are removed from the graph in real time), the model can instantly perceive the change in spatial topology without causing program crashes like with traditional fixed-dimensional matrix inputs. This greatly ensures the system's continuous operation capability in the face of sudden equipment failures.

[0064] Furthermore, after obtaining the action selection probability distribution of the deep reinforcement learning model for the current state, the system does not directly and blindly issue the action command with the highest probability, but instead performs a short-term forward-looking pre-deduction in combination with the digital twin environment.

[0065] Specifically, the system starts with the current real physical state and performs pre-simulation using Monte Carlo trajectory unfolding in a digital twin environment. Based on the action selection probability distribution output by the deep reinforcement learning model, the top K (e.g., K=3) candidate actions are selected as the first action for different trajectories. Subsequently, the system uses the digital twin environment to step forward and simulate the system's operating state within a finite time window (e.g., the next 3 minutes, encompassing several decision cycles), and uses a trained Critic network to evaluate the value of the future state. The system accumulates the expected compound reward sum for each simulated trajectory (denoted as ). ,in, (where t represents the time decay discount factor). Finally, the system compares the cumulative discounted rewards of each trajectory, selects the trajectory that maximizes the cumulative expected reward within the time window, and issues the optimal first step of that trajectory as the final scheduling instruction for the current cycle.

[0066] It's important to explain that, since deep reinforcement learning models are essentially black boxes based on statistical probability outputs, there's a very small probability that they will output a locally suboptimal solution. By introducing the aforementioned pre-deduction method, the system essentially performs a simulation before issuing commands. The advantage of this is that it can effectively identify and filter out short-sighted commands that seem to achieve high throughput in the short term but will cause the antenna to become stuck in a dead zone and unable to track high-priority tasks after a few minutes, thus selecting actions with true global optimal potential.

[0067] Furthermore, before outputting the final scheduling command to the physical antenna servo controller, the present invention also includes a security interception verification step.

[0068] Specifically, the system sends the derived scheduling command into a secure sandbox independent of the deep reinforcement learning model, comparing it against a set of pre-defined absolute hard constraint rules. These hard constraint rules are based on antenna physical limits and safe operating procedures, such as the antenna's physical rotation limit angle (to prevent excessive azimuth rotation that could cause internal RF cables to become entangled and break), operational safety red lines (e.g., limiting the antenna elevation angle to no less than 5 degrees to prevent high-power RF signals from radiating to people or buildings on the ground), and the transient peak current limit of the servo motor. The system verifies whether the scheduling command will trigger any of these red lines. If a command violates the rules, the system immediately intercepts the command and forcibly activates a predefined backup safety strategy (e.g., keeping the antenna pointing unchanged, or safely resetting the antenna to a safe zenith position at a low speed) to replace the command.

[0069] It should be noted that since the essence of inference lies in continuous exploration, the model's decisions are inherently unpredictable. In asset-heavy fields such as aerospace telemetry and control, where equipment safety requirements are extremely high, even the slightest erroneous instruction can cause expensive servo motors to malfunction. Therefore, this invention, through a security interception verification step, essentially sets up a deterministic firewall, separating and decoupling efficient optimization capabilities from the absolute security of industrial control, thus solving the problem of practical engineering application of reinforcement learning.

[0070] Furthermore, once the physical antenna executes a security-verified scheduling command, it generates a real physical interaction result. Based on this, the present invention also designs an online policy fine-tuning step, endowing the model with the ability to learn throughout its life.

[0071] Specifically, the system collects real-world interaction data, recording the actual state before execution, the actual scheduling instructions issued, the actual time overhead and link throughput (real rewards) incurred by the device, and the new real-world state after execution. This data, bearing the marks of actual physical device wear and tear, is then stored in an online experience replay pool. When the system is operating at night or during periods of low computational workload, this real-world feedback data is used to incrementally fine-tune the weights of the internal neural network of the already offline-trained deep reinforcement learning model with small learning steps.

[0072] It should be noted that as ground station equipment operates for many years, the gears of the antenna servo motor will wear down, and the rotational response speed will gradually slow down. The local microclimate electromagnetic environment may also undergo long-term, gradual changes. This means that the digital twin model in S1 will inevitably develop slight deviations from the real physical world after several months or years of operation. Through the aforementioned fine-tuning steps, the model can implicitly perceive the degradation of the equipment's mechanical performance through realistic reward feedback and adaptively adjust its scheduling strategy (e.g., the model will gradually learn to issue switching commands more early when facing aging antennas to compensate for their slower rotation speed). This not only prevents catastrophic forgetting in the model but also endows the system with the ability to evolve with the physical environment, extending the full lifecycle performance of the antenna scheduling system.

[0073] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for intelligent scheduling of satellite ground station antennas in a multi-satellite concurrent scenario, characterized in that, include: S1: Construct a digital twin scheduling environment, model the orbital dynamics of ground station antennas, multiple satellites, and the quality of communication links between antennas and satellites through the environment, and predict the dynamic changes of the quality of the communication links in the future time period; S2: Based on a deep reinforcement learning model, it is trained offline in the digital twin scheduling environment. Through simulated interaction with the environment, it learns a scheduling strategy that maximizes a multi-objective composite reward under preset constraints. S3: Using the trained deep reinforcement learning model, generate and output antenna scheduling instructions for the current decision cycle online based on the real-time perceived status of ground stations, satellites, and missions. 2.The intelligent scheduling method for satellite ground station antennas in multi-satellite concurrent scenarios according to claim 1, wherein, Modeling the ground station antenna includes: Construct a model of the dynamic response characteristics of an analog antenna electromechanical servo system; The model calculates the non-instantaneous time and trajectory of the antenna switching between different directions, and uses the time as one of the constraints for scheduling decisions. 3.The intelligent scheduling method for satellite ground station antennas in multi-satellite concurrent scenarios according to claim 1, characterized in that, Predicting the dynamic changes in the quality of the communication link over a future time period includes at least one of the following: The impact of future rainfall attenuation on different spatial pointing links is predicted based on meteorological data; Predict the probability of future radio frequency interference occurring in specific frequency bands and directions based on historical spectrum data.

4. The intelligent scheduling method for satellite ground station antennas in a multi-satellite concurrent scenario as described in claim 1, characterized in that, The S2 includes: The real-time sensed state information is constructed into a dynamic graph structure containing antenna nodes and satellite-mission nodes. The attributes of the nodes include the dynamic parameters of the antenna or satellite, and the edge attributes characterize the potential communication relationship between the antenna and the satellite. The deep reinforcement learning model understands the dependencies between different entities by processing the graph structure and makes decisions based on these dependencies.

5. The intelligent scheduling method for satellite ground station antennas in a multi-satellite concurrent scenario as described in claim 4, characterized in that, After the deep reinforcement learning model processes the graph structure, it outputs an action selection probability distribution for all visible and serviceable satellite-mission nodes of each idle antenna node in the graph structure, and determines a scheduling action for each antenna based on the probability distribution.

6. The intelligent scheduling method for satellite ground station antennas in a multi-satellite concurrent scenario as described in claim 1 or 4, characterized in that, The offline training adopts a course-based learning mechanism, including: By setting scheduling scenarios from simple to complex, the deep reinforcement learning model learns basic scheduling capabilities in simple scheduling scenarios and gradually transitions to complex scenarios with multiple constraints and dynamic environmental factors for policy optimization as the training process progresses.

7. The intelligent scheduling method for satellite ground station antennas in a multi-satellite concurrent scenario as described in claim 1, characterized in that, The multi-objective composite reward includes: The rewards are positively correlated with the total amount of data transmitted, with event rewards associated with completing high-priority tasks, load balancing rewards for balancing the load on each antenna, and handover penalties negatively correlated with antenna pointing switching overhead.

8. The intelligent scheduling method for satellite ground station antennas in a multi-satellite concurrent scenario as described in claim 1, characterized in that, The online generation of the antenna scheduling instructions includes: In each decision cycle, based on the current real state, a short-time domain pre-simulation is performed in the digital twin scheduling environment, and the optimal first step determined by the deep reinforcement learning model in the pre-simulation that maximizes the cumulative expected compound reward is selected as the final scheduling instruction for the current cycle.

9. The intelligent scheduling method for satellite ground station antennas in a multi-satellite concurrent scenario as described in claim 8, characterized in that, Before outputting the antenna scheduling command, a security verification step is also included, which includes: The final scheduling command is compared with a set of preset hard constraint rules, including physical rotation limits for the antenna and operational safety red lines. If the command violates the rules, a predefined backup strategy is activated to replace the command.

10. The intelligent scheduling method for satellite ground station antennas in a multi-satellite concurrent scenario as described in claim 1, characterized in that, The method further includes an online strategy fine-tuning step, which includes: During the operation of the method, real interaction data generated by the physical device executing scheduling instructions is collected, and the real interaction data is used to incrementally update the deep reinforcement learning model that has been trained offline.

Images