Imaging satellite task scheduling automatic modeling method based on large language model

Abstract

The invention discloses an imaging satellite task scheduling automatic modeling method based on a large language model, which automatically generates different types of optimization models and corresponding greedy algorithms through LLMs, not only can efficiently solve the scheduling problem of different scales of tasks, but also can flexibly adjust when processing complex constraints, and improves the task scheduling efficiency. The method shows strong automatic correction capability, and can realize superior resolving capability and adaptivity when processing large-scale tasks.

Description

Technical Field

[0001] The present invention belongs to the field of satellite mission planning and scheduling, and particularly relates to a method for automatically modeling the task scheduling of imaging satellites based on large language models. Background Art

[0002] Imaging satellites, as key tools for obtaining spatio-temporal information, are widely used in multiple fields such as resource exploration, disaster management, meteorological monitoring, environmental protection, and land use. Satellite imaging technology can not only provide accurate geographical information but also reflect global dynamic changes in real time, providing important support for scientific research and national security, etc. With the rapid development of remote sensing technology and the continuous expansion of application fields, the demand for high-resolution and high-timeliness remote sensing data also shows a rapid growth trend. Especially when dealing with complex tasks such as natural disasters, climate change, urbanization process, and environmental protection, it is particularly important to efficiently obtain remote sensing data. This makes the task scheduling problem of imaging satellites become more prominent and become a key problem in scientific research and technological applications.

[0003] The imaging satellite task scheduling problem (Imaging Satellite Task Scheduling, abbreviated as ISTS) is a widely recognized NP-hard problem, which involves how to reasonably arrange and schedule a large number of observation tasks in a multi-satellite system. Its core goal is to optimize the efficiency, quality, and timeliness of satellite imaging while following a series of complex constraints. To achieve this goal, task scheduling not only needs to consider how to reasonably allocate tasks to each satellite but also needs to comprehensively optimize in multiple aspects such as spatio-temporal resources, task priorities, satellite attitude adjustment, observation time, and energy consumption. Specifically, task scheduling needs to decide which tasks each satellite will execute, select appropriate observation time periods, coordinate task conflicts between satellites, and ensure that satellites can efficiently complete tasks within the specified time. In addition, factors such as the energy limit of the satellite system, the priority ranking between tasks, the time limit for satellite attitude adjustment, and the orbit limit of the satellite need to be incorporated into the decision-making process to achieve the optimal scheduling plan.

[0004] Despite the significant efforts made by existing research in this field, many methods based on models such as integer programming (IP), constraint satisfaction (CSP), vehicle routing problem (VRP), etc. have been proposed, and techniques such as tabu search, greedy algorithm, genetic algorithm, etc. are combined for solution, but still face multiple challenges. First, when solving large-scale and complex satellite mission scheduling problems, manual modeling and coding methods often have problems such as low efficiency and lack of flexibility. Manually constructing models is not only time-consuming and laborious, but also difficult to ensure the high efficiency and adaptability of the system when dealing with complex constraints and real-time data. Therefore, how to design more efficient and intelligent scheduling algorithms remains one of the important research directions at present. Second, poor dynamic adaptability is another major challenge. Since the time windows of satellite missions may change at any time, the mission scale may also fluctuate according to actual needs, and the mission priorities may be adjusted, the scheduling system needs to have a high degree of flexibility and adaptability. Adjusting the task arrangement in real time and ensuring efficient execution are the keys to improving scheduling efficiency. How to enable the scheduling algorithm to self-adjust according to real-time data and environmental changes is one of the problems that need to be solved urgently in current research. Finally, the existing multi-model collaboration methods are relatively lacking. Different models such as integer programming (IP) and vehicle routing problem (VRP) have their own advantages and applicable scenarios, but lack systematic comparative analysis and automated selection mechanisms. Therefore, how to integrate the advantages of different models and dynamically select the most suitable scheduling method according to the specific requirements of the mission is another important research direction for improving satellite mission scheduling efficiency.

[0005] In recent years, with the rapid development of large language models (LLMs) such as ChatGPT and GPT-4, these advanced artificial intelligence technologies have demonstrated powerful data processing and logical reasoning capabilities, providing new ideas for solving complex satellite mission scheduling problems. Through natural language processing technology, large language models exhibit unique advantages in handling complex problems, automated modeling, and algorithm generation. Their strong semantic understanding and reasoning capabilities enable large language models to automatically identify multiple constraint conditions in satellite missions (such as task time windows, priority rankings, satellite orbit restrictions, etc.) and generate appropriate scheduling plans through intelligent reasoning. Compared with traditional manual modeling methods, LLMs can handle complex constraints more efficiently and adjust scheduling strategies according to the real-time environment. Specifically, by leveraging LLMs for automated modeling and algorithm design, complex satellite mission scheduling scenarios can be efficiently addressed. Large language models can not only generate reasonable scheduling plans in a short time but also continuously optimize the model and adjust the scheduling plan through real-time data feedback. As the task scale expands and scheduling scenarios change continuously, LLMs can dynamically adjust scheduling strategies according to different environmental changes, thereby improving the efficiency and accuracy of task scheduling. In addition, LLMs can also support multi-strategy optimization, automatically selecting the most suitable solution method according to the specific requirements of the task through intelligent algorithms and adaptive mechanisms. Therefore, how to utilize LLMs to achieve efficient modeling, dynamic adjustment, and multi-strategy optimization of imaging satellite task scheduling problems (ISTS) has become an important direction of current technical research and the core challenge for the future development of satellite mission scheduling systems.

[0006] Generally speaking, the research on imaging satellite task scheduling problems faces huge challenges but also has broad application prospects. With the continuous development of remote sensing technology and artificial intelligence technology, how to ensure the efficiency of task scheduling while taking into account the flexibility, real-time nature, and multi-model collaboration of the system will be the key issues in future research and applications. And large language models provide new ideas and methods for solving these problems and are worthy of in-depth exploration and practice in the future field of satellite mission scheduling. Summary of the Invention

[0007] The present invention provides a data automated acquisition system to solve the problems of how to enable LLMs to automatically generate scheduling models adapted to different scenarios through a unified prompt template and how to enable LLMs to generate efficient algorithms matching the models through the prompt template, and at the same time, verify the adaptability of the algorithms through manual verification and iteratively improve the algorithms according to the optimized template.

[0008] To achieve the above object, the technical solution adopted by the present invention to solve the above technical problems is:

[0009] A method for automated modeling of imaging satellite mission scheduling based on large language models, which uses large language models (LLMs) to generate different models to process the imaging satellite task scheduling problem (ISTS). The imaging satellite task scheduling problem (ISTS) is a multi-objective optimization problem, including setting constraint conditions and objective functions;

[0010] Set constraint conditions, including uniqueness constraint, attitude adjustment time limit, observation duration limit, and time window limit, specifically including:

[0011]

[0012]

[0013] Δrt j ≥Δt j (3)

[0014] st j ≤T j ≤T j +Δt j ≤et j (4)

[0015] Equation (1) is the uniqueness constraint. To avoid resource waste, all tasks only need to be executed once by any one satellite. x ij is the decision variable. If the target has been assigned and completed, then x ij is 1, otherwise x ij is 0;

[0016] Equation (2) is the attitude adjustment time limit. After a task T j is completed after Δt j time, it takes an adjustment time of duration T to continue with the next task T j+1 ;

[0017] Equation (3) is the observation duration limit. The time for the satellite to observe each target task must be greater than the given imaging time of the target, otherwise the target cannot be considered completed;

[0018] Equation (4) is the time window limit. Each target task has a time window, and the entire execution time of the task must be within the visible time window to be considered a successful plan;

[0019] Set the objective function as:

[0020]

[0021] Among them, the variable Z is the maximum profit value, r j is the profit value of task j, x ij: 0-1 decision variable, where if task j is assigned to satellite i, then x ij = 1; otherwise, x ij = 0; T j is the start time of task j; Δt j is the theoretical imaging time of task j; st j is the start time of the time window of task j; et j is the end time of the time window of task j; Δrt j is the actual imaging time of task j; The objective function of ISTS is the weight of each satellite multiplied by the 0-1 decision variable, and the sum is the maximum satellite revenue value solved under this model;

[0022] The large language model LLMs generates five models of integer programming IP, constraint satisfaction problem CSP, vehicle routing problem VRP, knapsack problem KP, and job shop scheduling problem JSP based on a unified template. The generation process of these models depends on the input task scenarios and scheduling requirements. The input includes satellite system parameters, task information, time window constraints, task priorities, task revenues, and task dependencies into the models; LLMs can automatically derive optimization models applicable to different scheduling problems; during the model generation process, LLMs will formulate corresponding decision variables and constraint conditions according to the requirements of different optimization problems; these models are consistent in objective function and constraint conditions, but their decision variable designs are different.

[0023] Furthermore, the LLMs model also includes a greedy algorithm generation and iterative correction process; The greedy algorithm generation specifically includes: when dealing with different models, LLMs will automatically generate an appropriate greedy algorithm based on the specific characteristics and problem backgrounds of the models; LLMs analyzes the input model information and designs an algorithm that meets the actual requirements according to the greedy rules. For each model, LLMs will formulate a corresponding decision process in combination with priorities and constraint conditions; specifically including:

[0024] For the VRP model, LLMs generates a greedy algorithm based on task priority sorting. In this process, the system will give priority to selecting high-revenue tasks and dynamically adjust the satellite attitude or path planning during each decision. The priority of task selection is comprehensively judged according to revenue, time window, and resource constraints to ensure that each round of decision can maximize the overall revenue;

[0025] For the IP model, LLMs combine the characteristics of the model and adopt a targeted conflict detection mechanism. When generating algorithms, LLMs pay attention to the constraint conditions in the model, including the conflict between time windows and uniqueness constraints. LLMs will ensure through algorithm design that during the solution process, conflicts are avoided in task allocation. Through the conflict detection mechanism, violations of constraints can be monitored and corrected in real time to ensure that the solution output by the algorithm is legal and valid.

[0026] Furthermore, the iterative correction process includes: after initially generating the algorithm, the system enters the manual verification stage. The output results given by LLMs are carefully checked to verify the effectiveness and correctness of the algorithm. The goal of manual verification is to confirm whether the output of the algorithm fully meets all input constraint conditions, including time windows, resource limitations, and uniqueness constraints; through actual testing, evaluate whether the algorithm can correctly handle various boundary conditions and whether it can generate the optimal solution as expected;

[0027] If problems are found during manual verification, it enters the dynamic adjustment stage; if the conflict detection mechanism fails to accurately detect constraint conflicts, or in some cases the algorithm violates the uniqueness constraint, the optimization prompt template will be used to guide LLMs to regenerate the algorithm; during this process, specific error feedback is input into the prompt template, and by adjusting the parameters and descriptions in the template, it is ensured that LLMs can generate a corrected algorithm that meets the requirements; the iterative correction process will continue to loop until LLMs can output an efficient algorithm that fully meets all constraint conditions.

[0028] The imaging satellite task scheduling method based on large language models disclosed in the present invention can automatically generate different types of optimization models and their corresponding greedy algorithms through LLMs. It can not only efficiently solve the scheduling problems of tasks of different scales but also flexibly adjust when dealing with complex constraints, showing a strong automatic correction ability. When dealing with large-scale tasks, it can achieve superior solution capabilities and adaptability. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] In order to more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the following will briefly introduce the drawings required for the description of the specific embodiments or the prior art. Obviously, the drawings in the following description are some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0030] Figure 1 is the flowchart of the method of the present invention;

[0031] Figure 2 is the general prompt template of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0032] The following describes exemplary embodiments of the present application with reference to the accompanying drawings. Various details of the embodiments of the present application are included to facilitate understanding, and they should be considered merely exemplary. Therefore, those of ordinary skill in the art should recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of the present application. Similarly, descriptions of well-known functions and structures are omitted in the following description for clarity and conciseness.

[0033] As Figure 1 shown, a method for automated modeling of imaging satellite mission scheduling based on large language models uses large language models (LLMs) to generate different models to process the satellite mission scheduling problem (ISTS), where the satellite mission scheduling problem (ISTS) is a multi-objective optimization problem with an objective function and constraint variables.

[0034] Set constraint conditions, including uniqueness constraint, attitude adjustment time limit, observation duration limit, and time window limit, specifically including:

[0035]

[0036]

[0037] Δrt j ≥Δt j (3)

[0038] st j ≤T j ≤T j +Δt j ≤et j (4)

[0039] Equation (1) is the uniqueness constraint. To avoid resource waste, all tasks only need to be executed once by any one satellite. x ij is the decision variable. If the target has been assigned and completed, then x ij is 1, otherwise x ij is 0.

[0040] Equation (2) is the attitude adjustment time limit. After a task T j is completed after a time of Δt j , it takes an adjustment time of T to continue with the next task T j+1 .

[0041] Equation (3) is the observation duration limit. The time for the satellite to observe each target task must be greater than the given imaging time of the target, otherwise the target cannot be considered completed.

[0042] Equation (4) is the time window limit. Each target task has a time window, and the entire execution time of the task must be within the visible time window to be considered a successful plan;

[0043] Set the objective function as:

[0044]

[0045] where the variable Z is the maximum profit value, and r j is the profit value of task j, and x ij : a 0-1 decision variable, where if task j is assigned to satellite i, then x ij = 1; otherwise, x ij = 0; T j is the start time of task j; Δt j is the theoretical imaging time of task j; st j is the start time of the time window of task j; et j is the end time of the time window of task j; Δrt j is the actual imaging time of task j; The objective function of ISTS is the weight of each satellite multiplied by the 0-1 decision variable, and the sum is the maximum profit value of the satellite solved under this model;

[0046] The large language model LLMs generates five models of integer programming IP, constraint satisfaction problem CSP, vehicle routing problem VRP, knapsack problem KP, and job shop scheduling problem JSP based on a unified template. The generation process of these models depends on the input task scenarios and scheduling requirements. The input includes satellite system parameters, task information, time window constraints, task priorities, task profits, and task dependencies into the models; LLMs can automatically derive optimization models suitable for different scheduling problems; during the model generation process, LLMs will formulate corresponding decision variables and constraint conditions according to the requirements of different optimization problems; these models are consistent in objective function and constraint conditions, but their decision variable designs are different.

[0047] Furthermore, the LLMs model also includes a greedy algorithm generation and iterative correction process; The greedy algorithm generation specifically includes: when processing different models, LLMs will automatically generate an appropriate greedy algorithm based on the specific characteristics and problem background of the model; LLMs analyzes the input model information and designs an algorithm that meets the actual needs according to the greedy rules. For each model, LLMs will formulate a corresponding decision process in combination with priorities and constraint conditions; Specifically includes:

[0048] For the VRP model, LLMs generate a greedy algorithm sorted based on task priorities. In this process, the system preferentially selects high-reward tasks and dynamically adjusts satellite attitudes or path planning during each decision-making. The priority of task selection is comprehensively judged according to rewards, time windows, and resource constraints to ensure that the overall reward is maximized for each round of decision-making.

[0049] For the IP model, LLMs combine the characteristics of the model and adopt a targeted conflict detection mechanism. When generating the algorithm, it pays attention to the constraint conditions in the model, including the conflict between time windows and uniqueness constraints. LLMs will ensure that no conflicts occur during task allocation during the solution process through algorithm design. The conflict detection mechanism can monitor and correct situations violating constraints in real time to ensure that the solution output by the algorithm is legal and valid.

[0050] Furthermore, the iterative correction process includes: after initially generating the algorithm, the system enters the manual verification stage, where a detailed check is performed on the output results given by LLMs to verify the effectiveness and correctness of the algorithm. The goal of manual verification is to confirm whether the output of the algorithm fully meets all input constraint conditions, including time windows, resource limitations, and uniqueness constraints; through actual testing, evaluate whether the algorithm can correctly handle various boundary conditions and whether it can generate the optimal solution as expected.

[0051] If problems are found during manual verification, it enters the dynamic adjustment stage; if the conflict detection mechanism fails to accurately detect constraint conflicts, or if the algorithm violates the uniqueness constraint in some cases, the optimization prompt template will be used to guide LLMs to regenerate the algorithm; during this process, specific error feedback is input into the prompt template, and by adjusting the parameters and descriptions in the template, ensure that LLMs can generate a corrected algorithm that meets the requirements. The iterative correction process will continuously loop until LLMs can output an efficient algorithm that fully meets all constraint conditions.

[0052] Figure 2 The general prompt template mainly consists of three parts:

[0053] 1) Prompt for modeling, which mainly includes:

[0054] ① Problem description:

[0055] Describes that multiple satellites need to complete various tasks within a specific time window, and each task requires a certain imaging time. The goal is to optimize the order and time of satellite task execution to maximize the total reward.

[0056] ② Goal: Maximize the total income from completed tasks.

[0057] ③ Constraint conditions:

[0058] A. Observation Duration: Each task must be observed for its required imaging time to be considered complete.

[0059] B. Attitude Conversion Time: A 20 - second interval is required between two consecutive tasks performed by the same satellite.

[0060] C. Uniqueness: Each task can only be executed once.

[0061] D. Time Window: The imaging time of a task must be completely within its visible time window for successful planning.

[0062] ④ Data Interpretation:

[0063] satelliteName: The name of each satellite.

[0064] imageTime: Imaging time, in seconds.

[0065] startTime: The time when the time window starts.

[0066] endTime: The time when the time window ends.

[0067] targetName: Each task has a unique name.

[0068] priority: Task weight, one targetname corresponds to one priority.

[0069] ⑤ Modeling Requirements: As an expert in satellite remote sensing measurement and control modeling, please use one or more of the ideas in (IP / VRP / CSP / JSP / KP) to build a model according to the content sent to you.

[0070] 2) Algorithm Hints, which mainly include:

[0071] ① Algorithm Generation Hints:

[0072] When the solution result shows that the same task is executed multiple times, it needs to be further improved.

[0073] Or when the algorithm reports an error, it needs to be improved according to the error message.

[0074] Or when the solution shows that when two adjacent tasks are performed by the same satellite, the interval time does not reach the attitude conversion time, it needs to be further improved.

[0075] ② Iterative optimization hint:

[0076] All data has been stored in an Excel file named '...', which has six columns. The column names and data formats of each column are consistent with the example data. Please use the greedy strategy to solve the problem according to the established model.

[0077] 3) Iterative optimization hint (Prompt for improving)

[0078] According to the results of the preliminary algorithm solution, manual verification is required. The inputs include: the preliminary solution generated by the algorithm, the feedback data in actual applications, and the potential problems found during the testing process. Through the analysis of these results, manual verification will provide some possible errors or improvement points. The output is an updated model or algorithm after optimization, which will include corrected constraints, optimized objective functions, and redesigned solution algorithms. For example, some assumptions in the model may be corrected, or the heuristic strategies in the algorithm may be adjusted to improve the efficiency and accuracy of the solution. This process can go through multiple rounds of feedback and optimization until a precise and efficient solution is finally output.

[0079] LLMs generate five models of IP, CSP, VRP, KP, and JSP based on a unified template. Table 1 shows the decision variables of each model:

[0080] Table 1 Decision variable names of different models

[0081]

[0082] x s,w,t ∈ {0, 1} is a binary variable indicating whether satellite s executes task t within time window W;

[0083] y t ∈ {0, 1} is a binary variable indicating whether task t has been executed;

[0084] Satellite j ∈ {Satellite w , EndTime w - ImageTime t} represents the start time of the task;

[0085] Satellite t ∈ {Satellite1, Satellite2,...} represents the satellite executing the task;

[0086] x t,s ∈ {0, 1} indicates whether task t is executed by satellite s;

[0087] t start,ts represents the start time of task t;

[0088] t end,ts represents the end time of task t.

[0089] When dealing with different models, LLMs (Large Language Models) will automatically generate suitable greedy algorithms based on the specific characteristics of the models and the problem context. First, LLMs will analyze the input model information and design algorithms that meet the actual requirements according to the greedy rules. For each model, LLMs will combine specific priorities and constraints to formulate an efficient decision-making process.

[0090] For example, for the VRP (Vehicle Routing Problem) model, LLMs will generate a greedy algorithm based on task priority sorting. In this process, the system will preferentially select high-reward tasks and dynamically adjust the satellite attitude or path planning at each decision. The priority of task selection is comprehensively judged based on rewards, time windows, and resource constraints to ensure that each round of decision-making can maximize the overall reward. In this way, the algorithm can make optimal decisions in a complex constraint environment and quickly find feasible solutions.

[0091] For the IP (Integer Programming) model, LLMs will design a targeted conflict detection mechanism in combination with the characteristics of the model. When generating the algorithm, special attention is paid to the constraint conditions in the model, such as the conflict between time windows and uniqueness constraints. LLMs will ensure through algorithm design that during the solution process, conflicts are avoided during task allocation, especially in terms of time limits or resource uniqueness, to avoid violating constraints. These conflict detection mechanisms can monitor and correct any constraint violations in real time to ensure that the solutions output by the algorithm are legal and effective.

[0092] 2. Iterative correction process:

[0093] After the algorithm is initially generated, the system enters the manual verification stage. In this stage, we will conduct a detailed inspection of the output results given by LLMs to verify the effectiveness and correctness of the algorithm. Specifically, the goal of manual verification is to confirm whether the output of the algorithm fully meets all input constraint conditions, such as time windows, resource limitations, uniqueness constraints, etc. Through actual tests, we evaluate whether the algorithm can correctly handle various boundary cases and whether it can generate the optimal solutions as expected.

[0094] If problems are found during manual verification, the dynamic adjustment phase will begin. If errors are found in the model, such as the conflict detection mechanism failing to accurately detect constraint conflicts or the algorithm violating the uniqueness constraint in certain cases, we will optimize the prompt template to guide the LLMs to regenerate the algorithm. During this process, we will input specific error feedback (such as the details of the constraint conflict) into the prompt template and ensure that the LLMs can generate a corrected algorithm that meets the requirements by adjusting the parameters and descriptions in the template. For example, if it is found that the algorithm does not make appropriate adjustments when there is a time window conflict, we will clearly inform the LLMs how to incorporate a more precise conflict detection and handling mechanism into the algorithm.

[0095] This iterative correction process will continue to loop until the LLMs can output an efficient algorithm that fully meets all constraint conditions. In this way, the model and algorithm are optimized in each feedback, ultimately achieving an accurate and efficient solution. Through repeated manual verification and dynamic adjustment, we ensure the reliability and stability of the generated algorithm to adapt to more complex and diverse application scenarios.

[0096] The main objective of the experiment of the present invention is to verify the effectiveness of LLMs in automatically generating optimized models and algorithms and to evaluate their performance in solving the Imaging Satellite Task Scheduling (ISTS) problem. By generating different types of optimized models (such as IP, CSP, VRP, KP, and JSP) and the corresponding GAs, it aims to comprehensively evaluate the application potential of LLMs in automatic modeling, task scheduling, and optimization. At the same time, this experiment also compares the performance of LLMs in constructing different models under different task scales through testing different-scale task datasets, and reveals the advantages and disadvantages of various optimized models in solving the ISTS problem, thereby providing a theoretical basis for practical applications.

[0097] The computer has a hardware configuration of an Intel i7 12700H CPU, an NVIDIA GTX 3060 GPU, and 32 GB of RAM, the operating system is Windows 11, and the development environment uses Python 3.12. The LLMs tool used in the experiment is ChatGPT-4. The spatial range of the task scenario is from 74°E to 133°E longitude and from 3°N to 53°N latitude, and the time span is from 00:00 to 24:00 on April 20, 2013. The satellite system used is designed based on the Walker constellation, with a semi-major axis of the orbit of 7200 km, an eccentricity of 0.00067, and an inclination of 96.576°. These parameters ensure the real effectiveness and representativeness of the task scenario.

[0098] Five satellite imaging task datasets of different scales were designed in the experiment to simulate the actual scheduling requirements of tasks of different scales. The scale of the datasets gradually increases from the "50_1" dataset (indicating that 1 satellite executes 50 tasks) to the "500_6" dataset (indicating that 6 satellites execute 500 tasks). Each dataset includes important data such as satellite information, task information, time window information, imaging time, and task revenue information, and evaluates the performance and solution efficiency of different models in large-scale task scheduling by gradually increasing the task scale.

[0099] In the experiment, LLMs automatically generated a variety of optimization models based on a pre-designed general prompt template, including integer programming (IP), constraint satisfaction problem (CSP), vehicle routing problem (VRP), knapsack problem (KP), and job shop scheduling problem (JSP). The generation process of these models depends on the input task scenarios and scheduling requirements, and the input mainly includes basic parameters of the satellite system, task information, time window constraints, task priorities, task revenues, and task dependencies. Through these inputs, LLMs can automatically derive optimization models applicable to different scheduling problems. During the model generation process, LLMs will design specific decision variables and constraint conditions according to the requirements of different optimization problems. These models are consistent in objective function and constraint conditions, but their decision variable designs are different, mainly reflecting the emphasis of different models on task allocation, time arrangement, and task execution status.

[0100] 1) VRP model (Vehicle Routing Problem)

[0101] Input: The input of the VRP model not only includes the basic information of satellites and tasks, but also includes the time dependencies between tasks, the orbital constraints of satellites, and the status information of task execution. These input parameters ensure that the model can take into account the timeliness and rationality of task scheduling. Specifically, the input includes the following aspects:

[0102] ①: Satellite information: The number of satellites, the capabilities of each satellite, working time windows, orbital restrictions, etc.

[0103] ②: Task information: The basic information of each task, including the execution time of the task, the revenue of the task, the priority of the task, etc.

[0104] ③: Task time dependencies: Describe the sequential execution relationships between tasks (e.g., some tasks must start after other tasks are completed).

[0105] ④: Satellite orbital constraints: The availability and orbital restrictions of satellites at different time periods, which may involve the maximum flight time of satellites, the minimum flight interval, etc.

[0106] ⑤: Task execution status: Information such as whether the task has been assigned, the start time and end time of the task, etc., which is crucial for time window scheduling.

[0107] Design features: In the design of the VRP model, while assigning tasks, special attention is paid to the execution time of tasks and the dependencies between tasks. Its decision variables not only consider whether a task is assigned, but also need to ensure that the task is executed within an appropriate time window, and the start time and end time of the task must meet the constraints. In addition, the design of the VRP model also takes into account the orbital constraints of satellites and the interrelationships between tasks, and can effectively handle the time window constraints of tasks and the constraints on the execution order of tasks.

[0108] Specific mathematical model:

[0109] Equation (6): Represents the objective function, and the goal is to maximize the total revenue (or total profit), where p j is the revenue of task j, and x ij is the decision variable, indicating whether satellite i executes task j (1 means execution, 0 means non-execution).

[0110] Equation (7): Represents the observation duration constraint. Each task j must be observed for its required imaging time t image,i to be considered completed.

[0111] Equation (8): Represents the attitude conversion time constraint. There needs to be at least a 20-second interval between two consecutive tasks executed by the same satellite i.

[0112] Equation (9): Represents the uniqueness constraint. Each task j can only be executed once.

[0113] Equation (10): Represents the time window constraint. The imaging time of task j must be completely within its visible time window.

[0114] Equation (11): Represents the binary decision variable constraint. The decision variable x ij must be binary, that is, the task is either executed (1) or not executed (0).

[0115]

[0116]

[0117]

[0118]

[0119]

[0120]

[0121] The VRP model ensures the optimization of satellite mission scheduling by defining the objective function and a series of constraints, while meeting various requirements and limitations for task execution.

[0122] 2) IP model (Integer Programming Model)

[0123] Input: The input of the IP model mainly focuses on information such as task allocation requirements, task benefits, and task time windows. These inputs form the basis for model optimization and generally include the following aspects:

[0124] ①: Task allocation requirements: Each task must be assigned to a specific satellite and a specific time window.

[0125] ②: Task benefits: Each task may have different benefit values, and the goal is to maximize the total benefit through scheduling.

[0126] ③: Time window: Each task has a designated time window within which the task must be completed.

[0127] ④: Satellite information: The available time window and resource limitations of the satellite determine when the satellite can execute tasks.

[0128] The design of the IP model focuses on ensuring the uniqueness of tasks and the rationality of task allocation. Its decision variables mainly revolve around task allocation, ensuring that each task is assigned to only one specific satellite and time window. The objective function usually seeks the optimal solution by maximizing task benefits. Through the constraint conditions of integer programming, the IP model can precisely control the task allocation process, avoiding problems such as duplicate task allocation or unreasonable time arrangements.

[0129] Specific mathematical model:

[0130] Equation (12): The goal is to maximize the total benefit (or total profit), where priority t is the priority of task t, and y t is the decision variable indicating whether task t is executed (1 means executed, 0 means not executed).

[0131] Equation (13): Represents the observation duration constraint. The imaging time of each task t must be between the start time and the end time and satisfy the imaging duration

[0132] Equation (14): Represents the attitude conversion time constraint. There needs to be at least a 20 - second interval between two consecutive tasks executed by the same satellite.

[0133] Equation (15): Represents the uniqueness constraint. Each task t can only be assigned to one satellite.

[0134] Equation (16): Represents the time window constraint. The imaging time of task t must be completely within its visible time window.

[0135]

[0136]

[0137]

[0138]

[0139]

[0140] The IP model ensures the optimization of satellite task scheduling by defining the objective function and a series of constraint conditions, while meeting various requirements and limitations for task execution.

[0141] 3) KP model (Knapsack Problem Model)

[0142] Input: The input of the KP model is similar to that of the IP model, but it emphasizes more on the management of resource constraints. Its input includes:

[0143] ①: Task assignment requirements: Tasks need to be assigned to specific satellites and meet the time window requirements.

[0144] ②: Task benefits: Each task has a benefit value associated with it. Usually, the goal is to maximize the total benefit.

[0145] ③: Time window: The execution time of each task is limited within a specific time period.

[0146] ④: Resource constraints: In addition to the basic information of the tasks, the KP model usually also includes limitations on other resources, such as the maximum load of the satellite at different time periods, other physical resources required by the tasks, etc.

[0147] The decision variables of the KP model focus on the unique assignment of tasks, ensuring that each task can only be executed by one satellite within a specific time window. Different from the IP model, the KP model emphasizes more on the optimization of task assignment under limited resources. The KP model often adds additional resource limitations, such as time limitations, satellite orbit limitations, etc., to the objective function, which enables the model to handle more complex constraint conditions, such as the maximum load of the satellite.

[0148] Specific mathematical model:

[0149] Equation (17): The goal is to maximize the total benefit (or total value), where c j is the value of task j, and x ijkis a decision variable indicating whether satellite \(i\) executes task \(j\) within time window \(k\) (1 means execution, 0 means non - execution).

[0150] Equation (18): Represents the observation duration constraint. Each task \(k\) must be observed for its required imaging time to be considered completed.

[0151] Equation (19): Represents the attitude transition time constraint. There needs to be at least a 20 - second interval between two consecutive tasks executed by the same satellite \(i\) within time window \(k\).

[0152] Equation (20): Represents the uniqueness constraint. Each task \(j\) can only be executed once.

[0153] Equation (21): Represents the time window constraint. The execution time \(t\) of the task ijk must be within time window \(k\) and satisfy the imaging time.

[0154] Equation (22): Represents the binary decision variable constraint. The decision variable \(x\) ijk must be binary, that is, the task is either executed (1) or not executed (0).

[0155]

[0156]

[0157]

[0158]

[0159]

[0160]

[0161] The KP model ensures the optimization of satellite task scheduling by defining the objective function and a series of constraint conditions, while meeting various requirements and limitations for task execution.

[0162] 4) CSP model (Constraint Satisfaction Problem model)

[0163] Input: The input of the CSP model is relatively complex. In addition to basic task assignment and time windows, it also includes task priorities, task sequence requirements, and more detailed resource constraints. Specifically, the input of the CSP model includes:

[0164] ①: Task assignment information: Execution time of the task, benefit of the task, priority of the task, etc.

[0165] ②: Task sequence constraints: Requirements for the execution sequence between tasks. Some tasks must be executed before or after other tasks.

[0166] ③: Task Priority: The priority or importance of a task usually determines which tasks should be scheduled first.

[0167] ④: Time Window: The time window for each task ensures that the task is executed within a limited time.

[0168] ⑤: Resource Constraints: For example, the availability of satellites, the physical resources required for tasks, etc., may involve the scheduling of multiple resources.

[0169] The design of the CSP model emphasizes handling complex constraint conditions. It not only needs to consider task allocation but also the execution order and priority of tasks to ensure that multiple constraints in the task scheduling process are satisfied. The CSP model can handle the constraints of task order, priority, and time window simultaneously, so it can effectively solve the scheduling problem in a multi-constraint environment.

[0170] Specific mathematical model:

[0171] Equation (23): The goal is to maximize the total revenue (or total value), where c j is the value of task j, and x ijk is a decision variable indicating whether satellite i executes task j within time window k (1 means execution, 0 means non-execution).

[0172] Equation (24): Represents the observation duration constraint. Each task j must be observed for its required imaging time to be considered completed.

[0173] Equation (25): Represents the attitude conversion time constraint. There needs to be at least a 20-second interval between two consecutive tasks executed by the same satellite i within time window k.

[0174] Equation (26): Represents the uniqueness constraint. Each task j can only be executed once.

[0175] Equation (27): Represents the time window constraint. The execution time of a task must be within the time window and satisfy the imaging time.

[0176]

[0177]

[0178]

[0179]

[0180] startTime j ≥ startTime window and endTime j ≤ endTimewindow (27)

[0181] The goal of the CSP model is to find a solution that satisfies all the constraints, that is, each task is assigned to a suitable satellite and starts and ends at the appropriate time, while meeting the requirements of imaging time, time window, and attitude conversion. This model is very useful in solving complex scheduling and resource allocation problems because it can systematically handle and solve multiple interrelated constraints.

[0182] 5) JSP model (Job Shop Scheduling Problem model)

[0183] Input: The input of the JSP model is the most complex. In addition to basic information such as task assignment, time window, and task priority, it also includes more scheduling constraints. The input of the JSP model includes:

[0184] ①: Task information: including the execution time of the task, the revenue of the task, the priority of the task, the dependency relationship between tasks, etc.

[0185] ②: Task priority and sequence requirements: the execution sequence requirements of the tasks and the priority of each task.

[0186] ③: Satellite information: the available time window of the satellite, satellite resources, etc.

[0187] ④: Time window and resource constraints: the time window constraints of the tasks and the resource usage limitations of the satellite in each time period.

[0188] The JSP model mainly focuses on task scheduling and resource allocation. It not only needs to consider task priority and time window, but also needs to precisely handle the sequence requirements of tasks. Due to involving multiple tasks and multiple resources, the design of the JSP model can handle more complex scheduling constraints, such as resource allocation and task execution sequence. The JSP model is usually applied to scheduling problems with multiple tasks and complex constraints and can handle large-scale task assignment problems.

[0189] Specific mathematical model:

[0190] Equation (28): The goal is to maximize the total revenue (or total value), where c j is the value of task j, and x ijk is a decision variable indicating whether satellite i executes task j in time window k (1 means execute, 0 means not execute).

[0191] Equation (29): Represents the observation duration constraint. Each task j must be observed for its required imaging time to be considered completed.

[0192] Equation (30): Represents the attitude conversion time constraint. There needs to be an interval of at least 20 seconds between two consecutive tasks performed by the same satellite i within the time window k.

[0193] Equation (31): Represents the uniqueness constraint. Each task j can only be executed once.

[0194] Equation (32): Represents the time window constraint. The execution time of the task must be within the time window and satisfy the imaging time.

[0195]

[0196]

[0197]

[0198]

[0199] startTime j ≥startTime window and endTime j ≤endTime window (32)

[0200] The goal of the JSP model is to find a solution that satisfies all the constraints, that is, each task is assigned to a suitable satellite and starts and ends at the appropriate time, while meeting the requirements of imaging time, time window, and attitude conversion. This model is very useful in solving complex scheduling and resource allocation problems because it can systematically handle and solve multiple interrelated constraints.

[0201] After the optimization models are generated, the LLMs continue to automatically generate the corresponding greedy algorithms according to the prompt templates for each optimization model. The core idea of the algorithm is to sort the benefits of the tasks and assign the tasks with the maximum benefits within the time window of each satellite to achieve the optimization of task scheduling. The algorithm strictly follows the constraints when assigning tasks to ensure the effectiveness and legality of task scheduling. The specific characteristics of each model are as follows:

[0202] 1) Greedy algorithm for the VRP model

[0203] The greedy algorithm for the VRP model focuses on task assignment and satellite orbit constraints. The algorithm sorts the benefits of the tasks and selects the tasks with the maximum benefits within the time window of the satellite for scheduling. Specifically, the algorithm will evaluate each task one by one and preferentially assign the task with the highest benefit based on the benefit of the task, time window limit, and satellite availability.

[0204] ① Orbit and attitude constraints: During the initial generation, due to insufficient consideration of the satellite's attitude adjustment constraints, the scheduling results were not entirely reasonable. After correction, the algorithm can take into account the constraints of the satellite orbit and optimize the task scheduling to ensure that tasks can be reasonably arranged within the available time.

[0205] ② Time window management: When allocating tasks, strictly adhere to the time window to ensure that tasks do not exceed the specified time interval.

[0206] 2. Greedy algorithm for the IP model

[0207] The greedy algorithm for the IP model depends on the benefits of tasks to determine the task allocation order. First, sort all tasks by benefit value and give priority to allocating tasks with high benefits. The algorithm allocates tasks to the most suitable satellite and time window by considering the time window of each task and the availability of the satellite.

[0208] ① Benefit maximization: The goal of the IP model is to maximize task benefits. Therefore, the core of the greedy algorithm is to select tasks with the largest benefit value for allocation.

[0209] ② Time window problem: During the initial generation, the time window format was not accurately understood, resulting in the algorithm not being able to run properly. However, through adjustment, the algorithm can correctly understand the time window data format and successfully optimize task scheduling.

[0210] Unique task allocation: Each task is allocated to a unique satellite and time window to avoid duplicate scheduling.

[0211] 3. Greedy algorithm for the KP model

[0212] The greedy algorithm for the KP model is similar to the IP model but emphasizes more on resource constraint problems. The algorithm sorts tasks according to the benefit value of tasks and resource constraints (such as satellite availability, time limit, etc.), and then allocates tasks according to the priority. Each time, select the task with the largest current benefit to ensure that resource constraints are not violated.

[0213] ① Resource constraint priority: Different from the IP model, the KP model pays more attention to the resource constraints of satellites and time periods. When allocating tasks each time, the algorithm will specifically consider these resource constraints to ensure that task allocation meets the maximum availability of resources.

[0214] ② Uniqueness of task allocation: Ensure that each task can only be allocated to one satellite and one time window to avoid duplicate task allocation.

[0215] 4. Greedy algorithm for the CSP model

[0216] The greedy algorithm of the CSP model not only needs to consider the benefits of tasks, but also must consider task order, priority, and resource constraints. The algorithm first sorts tasks according to their benefits, and then gradually assigns tasks to appropriate satellites and time windows according to the priority and order requirements of the tasks.

[0217] ① Task priority and order constraints: The CSP model attaches particular importance to task priority and order constraints. When the greedy algorithm assigns tasks, it will give priority to handling high-priority tasks and ensure that the order requirements of tasks are met.

[0218] ② Constraint correction: Initially, the greedy algorithm of the CSP model did not fully follow the uniqueness constraint, resulting in tasks possibly being assigned repeatedly. However, after adjustment, the algorithm can ensure task uniqueness and avoid the situation of repeated scheduling.

[0219] 5. The greedy algorithm of the JSP model

[0220] The greedy algorithm of the JSP model is optimized for a multi-task, multi-resource environment. Its core idea is to sort tasks according to task priority and benefit value, and select the optimal task for assignment under multiple resource constraints. The algorithm will evaluate the execution order, time window, priority, and resource availability of each task, and select the optimal task one by one.

[0221] ① Resource and order priority: The JSP model not only focuses on the benefits of tasks, but also needs to comprehensively consider the execution order of tasks, the usage of resources, and the priority of tasks.

[0222] ② Scheduling order correction: Initially, the greedy algorithm of the JSP model can successfully complete task scheduling. However, when faced with more complex constraints (such as resource competition and task dependencies), the algorithm may need to adjust the order of task assignment and resource arrangement.

[0223] During the algorithm generation process, LLMs demonstrated their excellent adaptability and adjustment capabilities. For example, when initially generating the greedy algorithms of the JSP and KP models, no errors occurred and task scheduling was successfully completed. When the IP model was initially generated, the algorithm could not run properly because the time window data format was not correctly understood. However, after one prompt, it was successfully corrected and task scheduling was completed. The greedy algorithm of the CSP model violated the uniqueness constraint initially, but was also fixed after one prompt. When the VRP model was initially generated, the scheduling was not completely reasonable because the attitude adjustment constraint was not fully considered. However, after two prompts, the problem was finally successfully solved. Through these adjustments, LLMs demonstrated their flexibility and tuning capabilities in complex scheduling problems, ensuring the accuracy and efficiency of the generated algorithms.

[0224] Table 2 Benefit values of different models in different scale scenarios

[0225]

[0226] The experimental results show that LLMs perform excellently in automatically generating optimized models and algorithms, especially for the VRP model, demonstrating strong adaptability and computational efficiency, and being able to provide effective solutions for task scheduling problems of different scales.

[0227] The experiments of the present invention verify the significant advantages of LLMs in automatically generating optimized models and algorithms for the imaging satellite task scheduling (ISTS) problem. By automatically generating different types of optimized models (IP, CSP, VRP, KP, JSP) and their corresponding greedy algorithms, LLMs can not only efficiently solve the scheduling problems of tasks of different scales, but also flexibly adjust when dealing with complex constraints, showing strong automatic correction ability. Especially when dealing with large-scale tasks, its superior solving ability and adaptability are demonstrated.

[0228] The above specific embodiments do not constitute a limitation on the protection scope of the present application. Those skilled in the art should understand that various modifications, combinations, sub-combinations and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions and improvements made within the spirit and principle of the present application shall be included within the protection scope of the present application.

Claims

1. A method for automated modeling of imaging satellite mission scheduling based on a large language model, characterized in that: Large language models (LLMs) are used to generate different models to process the satellite mission scheduling problem (ISTS), wherein the satellite mission scheduling problem (ISTS) is a multi-objective optimization problem, including setting constraints and objective functions; Set constraints, including uniqueness constraints, attitude adjustment time limits, observation duration limits, and time window limits, including: Δrt j ≥Δt j (3) st j ≤T j ≤T j +Δt j ≤et j (4) Formula (1) is a unique constraint. To avoid wasting resources, all tasks only need to be executed once by any satellite. ij is a decision variable. If the goal has been assigned and completed, then x ij is 1, otherwise x ij is 0; Formula (2) is the time limit for posture adjustment. j After Δt j After the time is completed, an adjustment time of T is required before the next task T can be continued j+1 ; Formula (3) is the observation duration limit. The satellite observation time for each target task must be greater than the given imaging time of the target, otherwise the target cannot be considered completed. Formula (4) is the time window constraint. Each target task has a time window. The entire execution time of the task must be within the visible time window to be considered a successful plan. Set the objective function to: Among them, variable Z is the maximum profit value, r j is the profit value of task j, x ij : 0-1 decision variable, where if task j is assigned to satellite i, then x ij =1; otherwise, x ij =0; T j is the start time of task j; Δt j is the theoretical imaging time of task j; st j is the start time of the time window of task j; et j is the end time of the time window of task j; Δrt j is the actual imaging time of task j; the objective function of ISTS is the weight of each satellite multiplied by a 0-1 decision variable, and the sum is the maximum benefit value of the satellite solved under the model; The large language model LLMs generates five models, namely integer programming IP, constraint satisfaction problem CSP, vehicle path planning VRP, knapsack problem KP, and job shop scheduling problem JSP, based on a unified template. The generation process of these models depends on the input task scenarios and scheduling requirements. The input includes satellite system parameters, task information, time window constraints, task priorities, task benefits, and task dependencies into the model; LLMs can automatically derive optimization models suitable for different scheduling problems; in the process of model generation, LLMs will formulate corresponding decision variables and constraints according to the requirements of different optimization problems. These models are consistent in objective functions and constraints, but their decision variables are different.

2. The method according to claim 1, characterized in that: The LLMs model also includes a greedy algorithm generation and iterative correction process; the greedy algorithm generation specifically includes: when processing different models, LLMs will automatically generate an adaptive greedy algorithm based on the specific characteristics of the model and the problem background; LLMs analyzes the input model information and designs an algorithm that meets actual needs based on greedy rules. For each model, LLMs will combine priorities and constraints to develop a corresponding decision-making process; specifically including: For the VRP model, LLMs generates a greedy algorithm based on task priority sorting. The system will give priority to high-benefit tasks and dynamically adjust the satellite attitude or path planning at each decision. The priority of task selection is comprehensively judged based on benefit, time window and resource constraints to ensure that each round of decision-making can maximize the overall benefit. For IP models, LLMs combines the characteristics of the model and adopts a targeted conflict detection mechanism. When generating the algorithm, it pays attention to the constraints in the model, including the conflict between the time window and the uniqueness constraint, to ensure that conflicts are avoided when allocating tasks during the solution process. The conflict detection mechanism can monitor and correct the violation of constraints in real time. For the KP model, the greedy algorithm generated by LLMs sorts the tasks according to their revenue values ​​and resource constraints, and then assigns tasks according to their priorities, selecting the task with the largest current revenue each time to ensure that resource constraints are not violated; For the CSP model, the greedy algorithm generated by LLMs not only considers the benefits of the tasks, but also the task order, priority, and resource constraints. The algorithm first sorts the tasks according to their benefits, and then gradually allocates the tasks to appropriate satellites and time windows according to the priority and order requirements of the tasks; For the JSP model, the greedy algorithm generated by LLMs is optimized for multi-task, multi-resource environments. It sorts tasks according to their priorities and benefit values, and selects the optimal tasks for allocation under multiple resource constraints. The algorithm evaluates the execution order, time window, priority, and resource availability of each task, and selects the optimal tasks one by one.

3. The method according to claim 2, characterized in that The iterative correction process includes: after the algorithm is initially generated, the system enters the manual verification stage, and the output results given by LLMs are checked in detail to verify the effectiveness and correctness of the algorithm. The goal of manual verification is to confirm whether the output of the algorithm fully meets all input constraints, including time windows, resource constraints, and uniqueness constraints; through actual testing, it is evaluated whether the algorithm can correctly handle various boundary conditions and whether it can generate the expected optimal solution; If problems are found during manual verification, the dynamic adjustment phase will be entered; if the conflict detection mechanism fails to accurately detect constraint conflicts, or in some cases the algorithm violates the uniqueness constraint, the LLMs will be guided to regenerate the algorithm through optimization of the prompt template; in this process, specific error feedback will be input into the prompt template, and by adjusting the parameters and descriptions in the template, it is ensured that LLMs can generate a corrected algorithm that meets the requirements; the iterative correction process will continue to cycle until LLMs can output an efficient algorithm that fully meets all constraints.

Images