A remote sensing satellite swarm adaptive task planning method and system based on a large language model

Abstract

The application relates to a remote sensing satellite group adaptive task planning method and system based on a large language model, which comprises the following steps: constructing an optimization model based on decision variables, constraint conditions and optimization objectives of remote sensing satellite group task planning; generating an initial task planning strategy by using a greedy algorithm based on the optimization model; iteratively optimizing the initial task planning strategy by using a large language model, generating a new round of task planning strategy each time, and scoring by taking the optimization objective as an evaluation function until a preset number of iteration optimization times is reached, stopping optimization, and outputting a task planning strategy with the highest score; and performing remote sensing satellite group adaptive task planning based on the task planning strategy with the highest score. The application can effectively perform multi-satellite distributed task planning on dynamic tasks, improve task benefits and reduce satellite imaging time consumption.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of remote sensing satellite cluster mission planning, in particular to a remote sensing satellite cluster adaptive mission planning method and system based on a large language model. BACKGROUND

[0002] A satellite cluster refers to a distributed satellite system composed of several satellites, which are usually independent of each other in orbit and do not perform external orbit adjustment control. Remote sensing satellites have been widely used in land resource investigation, disaster monitoring, and military intelligence acquisition due to their spatial position advantage. As the demand for spatial information increases, the capacity of a single satellite is limited and cannot meet the increasing demand of users, so major spacefaring countries and research institutions have launched satellite clusters. Remote sensing satellite cluster dynamic mission planning is a typical NP-hard problem, and it is difficult to find the optimal solution in a limited time, so the focus of mission planning is to find a high-quality feasible solution in a limited time.

[0003] Currently, research on remote sensing satellite cluster mission planning for dynamic tasks mainly includes centralized mission planning and distributed mission planning. The centralized method models the multi-satellite mission planning problem from a global perspective, mainly using heuristic algorithms, meta-heuristic algorithms and machine learning methods to solve the problem, which has the problems of poor robustness and slow response time, and the failure of the master node may cause the entire mission planning to fail.

[0004] The distributed mission planning method has better robustness and can fully utilize satellite computing resources to enhance the response capability to dynamic tasks. However, the existing distributed mission planning method has the problems of high algorithm complexity and insufficient generalization ability, and it is difficult to solve the problem of exponential growth of solution space size caused by the increase in the number of satellites. SUMMARY

[0005] The purpose of the present application is to provide a remote sensing satellite cluster adaptive mission planning method and system based on a large language model, which aims to solve the problem of remote sensing satellite cluster dynamic mission planning that adapts to the growth of the number of satellites. The invention uses a large language model to centrally optimize the mission planning algorithm, and each satellite uses a distributed decision-making method based on the optimized planning algorithm to schedule dynamic tasks. This method does not rely on human experience and can adaptively learn mission planning strategies to improve task revenue and resource utilization, and can be effectively extended to large-scale satellite scenarios.

[0006] To achieve the above purpose, the present application provides the following scheme:

[0007] A remote sensing satellite cluster adaptive mission planning method based on a large language model, comprising:

[0008] An optimization model is constructed based on decision variables, constraint conditions and optimization objectives of the remote sensing satellite group task planning;

[0009] An initial task planning strategy is generated based on the optimization model using a greedy algorithm;

[0010] The initial task planning strategy is iteratively optimized by a large language model, a new round of task planning strategy is generated each time, and the optimization objective is used as an evaluation function to score until a preset number of iterations is reached, the optimization is stopped, and the task planning strategy with the highest score is output;

[0011] Based on the task planning strategy with the highest score, the remote sensing satellite group adaptive task planning is performed.

[0012] Optionally, the decision variable is a decision state of a satellite j on a task i; wherein the decision state includes rejecting the task and accepting the task;

[0013] The constraint condition includes a maximum imaging time of a single orbit of a satellite, a minimum imaging time of a single imaging of a satellite, an imaging of a target by a satellite needs to be completed within a satellite visible window, and a time interval between adjacent two imaging tasks needs to be greater than a satellite attitude maneuvering time;

[0014] The optimization objective includes maximizing task revenue and maximizing residual available imaging time of a satellite.

[0015] Optionally, the optimization model is:

[0016]

[0017] wherein G represents an optimization objective function, a represents a weight factor, R(x) represents a total revenue of a planned task within a task planning period accounts for a maximum revenue proportion, T(x) represents a proportion of residual available imaging time of all satellites within a task planning period accounts for a maximum imaging time, x ij represents a decision state of a satellite j on a task i, m represents a total number of tasks, n represents a total number of satellites, t image represents a minimum imaging time of a single imaging of a satellite, t total represents a maximum imaging time of a single orbit of a satellite, represents a start time of a kth task accepted by a satellite j, represents an end time of the kth task accepted by the satellite j, t trans represents a satellite attitude maneuvering time.

[0018] Optionally, scoring with the optimization objective as an evaluation function includes:

[0019] The evaluation function is used to evaluate a new round of task planning strategy, and the evaluation score and the strategy code are stored in the program database together; wherein, the strategy code refers to a runnable Python program generated by the large model according to the prompt word, and the prompt word is a string text information artificially designed, describing the requirements of the task as the input of the large model;

[0020] The program database is updated based on the island model and the genetic algorithm, and the prompt word input is sampled and generated to the large language model; wherein, the island model divides the program database into multiple independent sub-populations, each sub-population evolves programs independently, and is reset periodically; the genetic algorithm selects the best performing program in the sub-population, and generates new program individuals through crossover and mutation, and the new program is tested by the evaluator, and if it meets the requirements, it is stored in the program database.

[0021] A remote sensing satellite swarm adaptive task planning system based on a large language model, the system comprising: a construction module, a strategy generation module, a task optimization module and an adaptive task planning module;

[0022] The construction module is configured to construct an optimization model based on decision variables, constraint conditions and optimization objectives of remote sensing satellite swarm task planning;

[0023] The strategy generation module is configured to generate an initial task planning strategy based on the optimization model using a greedy algorithm;

[0024] The task optimization module is configured to iteratively optimize the initial task planning strategy through a large language model, generate a new round of task planning strategy each time, and score the optimization objective as an evaluation function until a preset number of iteration optimization times is reached, stop optimization, and output the task planning strategy with the highest score;

[0025] The adaptive task planning module is configured to perform adaptive task planning of the remote sensing satellite swarm based on the task planning strategy with the highest score.

[0026] Optionally, the decision variable is a decision state of satellite j on task i; wherein, the decision state includes: rejecting the task and accepting the task;

[0027] The constraint condition includes: satellite single-track maximum imaging time, satellite single-imaging minimum time, satellite imaging on a target needs to be completed within a satellite visible window, and the time interval between adjacent two imaging tasks needs to be greater than the satellite attitude maneuvering time;

[0028] The optimization objective includes: maximizing task revenue and maximizing satellite remaining available imaging time.

[0029] Optionally, the optimization model is:

[0030]

[0031] wherein, G represents an optimization objective function, a represents a weight factor, R(x) represents a proportion of total income of a planning task in a planning period to maximum income, T(x) represents a proportion of all remaining available imaging time of a satellite in a planning period to maximum imaging time, x ij represents a decision state of satellite j to task i, m represents a total number of tasks, n represents a total number of satellites, t image represents a minimum imaging time of a satellite, total represents a maximum imaging time of a satellite in a single track, represents a start time of a kth task accepted by satellite j, represents an end time of a kth task accepted by satellite j, t trans represents a time of satellite attitude maneuvering.

[0032] Optionally, the task optimization module scores with an optimization objective as an evaluation function include:

[0033] The evaluation function is used to evaluate a new round of task planning strategy, and the evaluation score and the strategy code are stored in the program database together; wherein, the strategy code is a runnable Python program generated by the large model according to the prompt word, and the prompt word is a string text information artificially designed, describing the demand of the task as the input of the large model;

[0034] The program database is updated based on the island model and the genetic algorithm, and the prompt word input is generated by sampling and input to the large language model; wherein, the island model divides the program database into multiple independent sub-populations, each sub-population evolves the program independently, and is reset periodically; the genetic algorithm selects the best performing program in the sub-population, and generates new program individuals through crossover and mutation, and the new program is tested by the evaluator, and if it meets the requirements, it is stored in the program database.

[0035] The beneficial effects of the present application are:

[0036] The present application firstly constructs an optimization model based on the decision variables, constraint conditions and optimization objectives of the remote sensing satellite group task planning; secondly, generates an initial task planning strategy based on the optimization model using a greedy algorithm; then iteratively optimizes the initial task planning strategy through a large language model, generates a new round of task planning strategy each time and scores with an optimization objective as an evaluation function, until a preset number of iteration optimization is reached, stops optimization and outputs the task planning strategy with the highest score; finally, performs adaptive task planning of the remote sensing satellite group based on the task planning strategy with the highest score. The present application can effectively perform multi-satellite distributed task planning on dynamic tasks, improve task income and reduce satellite imaging time consumption.

[0037] The traditional centralized task planning method has high algorithm complexity, and it is difficult to respond to the demand of dynamic tasks in real time due to the limitation of satellite-ground communication time, and the distributed task planning method has lower time complexity and faster response speed for dynamic tasks.

[0038] Through simulation examples, it is verified that the method can be effectively extended to the scene of large-scale satellites, has stronger generalization ability than the traditional method, and has better planning results. BRIEF DESCRIPTION OF DRAWINGS

[0039] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed in the embodiments will be briefly introduced as follows. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0040] Figure 1 The task planning algorithm based on large language model of the embodiment of the present application is outlined in the graph.

[0041] Figure 2 The target function value iteration optimization curve of the embodiment of the present application is shown in the graph. DETAILED DESCRIPTION

[0042] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some embodiments of the present application, not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0043] In order to make the above-mentioned purposes, features and advantages of the present application more obvious and easy to understand, the present application will be further described in detail below with reference to the drawings and specific embodiments.

[0044] As shown in Figure 1 The embodiment proposes a remote sensing satellite group adaptive task planning method based on large language model, which includes:

[0045] Based on the decision variables, constraint conditions and optimization objectives of the remote sensing satellite group task planning, an optimization model is constructed;

[0046] Based on the optimization model, a greedy algorithm is used to generate an initial task planning strategy;

[0047] The initial task planning strategy is iteratively optimized by a large language model, a new round of task planning strategy is generated each time, and the optimization target is used as an evaluation function to score until a preset number of iterations is reached, the optimization is stopped, and the task planning strategy with the highest score is output.

[0048] Based on the task planning strategy with the highest score, the remote sensing satellite group is adaptively planned.

[0049] Further, the decision variable is: the decision state of satellite j on task i; wherein the decision state includes: rejecting the task and accepting the task;

[0050] The constraint condition includes: the maximum imaging time of a single satellite track, the minimum imaging time of a single satellite, the imaging of a satellite on a target must be completed within the satellite visible window, and the time interval between adjacent two imaging tasks needs to be greater than the satellite attitude maneuvering time;

[0051] The optimization target includes: maximizing the task benefit and maximizing the remaining available imaging time of the satellite.

[0052] Specifically, in this embodiment, the decision variable: x ij represents the decision state of satellite j on task i;

[0053]

[0054] Constraint condition:

[0055] The maximum imaging time of a single satellite track t total = 150s.

[0056] The minimum imaging time of a single satellite t image = 10s.

[0057] The imaging of a satellite on a target must be completed within the satellite visible window.

[0058] The time interval between adjacent two imaging tasks needs to be greater than the satellite attitude maneuvering time t trans = 30s.

[0059] Optimization target:

[0060]

[0061] z i = max (x i1 ,x i2 ,…,x in ) (5)

[0062] Wherein, G represents the optimization objective function, taking maximizing the task revenue and maximizing the satellite remaining available imaging time as the optimization objective. R(x) is the task revenue ratio, expressed by the total revenue of the planned task in the task planning period divided by the maximum revenue. T(x) is the average remaining available imaging time ratio, expressed by the total remaining available imaging time of all satellites in the task planning period divided by the maximum imaging time. a is the weight factor, m is the total number of tasks, and n is the total number of satellites.

[0063] The mathematical description of the research problem is determined, the decision variables, constraint conditions and optimization objectives of the optimization problem are determined, and the optimization model is constructed as shown in formula (6):

[0064]

[0065] Wherein, G represents the optimization objective function, a represents the weight factor, R(x) represents the total revenue of the planned task in the task planning period divided by the maximum revenue, T(x) represents the total remaining available imaging time of all satellites in the task planning period divided by the maximum imaging time, x ij represents the decision state of satellite j to task i, m represents the total number of tasks, n represents the total number of satellites, t image represents the minimum imaging time of a satellite, total represents the maximum imaging time of a satellite, represents the start time of the kth task accepted by satellite j, represents the end time of the kth task accepted by satellite j, t trans represents the satellite attitude maneuvering time.

[0066] Further, taking the optimization objective as the evaluation function to score includes:

[0067] The evaluation function is used to evaluate a new round of task planning strategy, and the evaluation score and the strategy code are stored in the program database together. The strategy code refers to the executable Python program generated by the large model according to the prompt word. The program database stores these strategy codes and their corresponding evaluation scores. The prompt word is a string of text information designed by humans, describing the requirements of the task, as input for the large model.

[0068] The program database is updated based on an island model and a genetic algorithm, and sampling generates prompt words input to the large language model. The island model divides the program database into multiple independent sub-populations (islands), each island independently evolves programs, and periodically resets to maintain diversity and avoid local optimization. The genetic algorithm selects the best performing program within the island, and generates new program individuals through crossover and mutation. The new program is tested by the evaluator, and if it meets the requirements, it is stored in the program database. Excellent programs are used for sampling to generate new programs, promoting continuous optimization and improvement, and ultimately achieving the goal of finding the best solution.

[0069] In this embodiment, the task planning algorithm based on the large language model is outlined as shown in Figure 1 The first step is to construct a distributed task planning program skeleton based on the greedy algorithm and an evaluation function based on the established optimization model, which are used to solve the optimization model and evaluate the score, respectively. Second, combined with the gpt-3.5-turbo pre-trained large language model and the evaluation function, the key program logic part, i.e. the greedy strategy with low initial score, is iteratively optimized on the initial program skeleton to obtain a strategy with higher score in the form of a program. The specific implementation steps of this embodiment are as follows:

[0070] 1. Determine the mathematical description of the research problem, determine the decision variables, constraint conditions and optimization objectives of the optimization problem, and construct the optimization model as shown in formula (6).

[0071] 2. Determine the simulation example and model parameters, use STK to build a simulation platform to construct a simulation scene, extract the time window when the satellite is visible to the target, and complete data preprocessing.

[0072] 3. Construct a distributed task planning program skeleton based on the earliest and most optimal greedy strategy of visible time, and use the objective function of the optimization model as the evaluation function.

[0073] 4. Set a fixed number of optimization iterations, use the gpt-3.5-turbo pre-trained large language model as the optimizer to iteratively optimize the initial greedy strategy, and generate a new round of task planning strategy.

[0074] 5. The evaluation function evaluates the new round of task planning strategy, and stores the evaluation score and strategy code together in the program database.

[0075] 6. Sample k programs and corresponding scores in the program database to construct prompt words, and pass them to the large language model to generate the next round of task planning strategy.

[0076] 7. Determine whether the number of iterative optimization has reached the set value, if not, return to step 4 and continue optimization. Otherwise, stop optimization and output the task planning strategy with the highest final score.

[0077] 8. Replace the greedy strategy in the obtained program skeleton with the policy code, calculate performance indicators such as the objective function, task revenue, and satellite residual available energy.

[0078] The technical focus of the embodiment is as follows:

[0079] A distributed task planning program skeleton based on a greedy algorithm and an evaluation function are constructed.

[0080] A program database is constructed in combination with a genetic algorithm, the program database is updated and optimized through an island model, and prompt words for inputting a pre-trained large language model are sampled.

[0081] The pre-trained large language model is used to iteratively optimize the key decision-making part, i.e., the greedy strategy, in the program skeleton.

[0082] The embodiment also proposes a remote sensing satellite swarm adaptive task planning system based on a large language model, which includes a construction module, a strategy generation module, a task optimization module, and an adaptive task planning module.

[0083] The construction module is used to construct an optimization model based on decision variables, constraint conditions, and optimization objectives of remote sensing satellite swarm task planning.

[0084] The strategy generation module is used to generate an initial task planning strategy based on the optimization model using a greedy algorithm.

[0085] The task optimization module is used to iteratively optimize the initial task planning strategy through a large language model, generate a new round of task planning strategy each time, and score it using the optimization objective as an evaluation function until a preset number of iterations is reached, stop optimization, and output the task planning strategy with the highest score.

[0086] The adaptive task planning module is used to perform adaptive task planning of the remote sensing satellite swarm based on the task planning strategy with the highest score.

[0087] Further, the decision variable is the decision state of satellite j on task i, and the decision state includes rejecting the task and accepting the task.

[0088] The constraint conditions include the maximum imaging time of a satellite on a single track, the minimum imaging time of a satellite on a single track, the imaging of a satellite on a target needing to be completed within the satellite's visible window, and the time interval between adjacent two imaging tasks needing to be greater than the satellite's attitude maneuvering time.

[0089] The optimization objectives include maximizing the task revenue and maximizing the satellite's residual available imaging time.

[0090] Further, the optimization model is:

[0091]

[0092] wherein G represents an optimization objective function, a represents a weight factor, R(x) represents a proportion of total revenue of planning tasks in a task planning period to maximum revenue, T(x) represents a proportion of all satellites' remaining available imaging time in a task planning period to maximum imaging time, x ij represents a decision state of satellite j on task i, m represents a total number of tasks, n represents a total number of satellites, t image represents a minimum imaging time of a satellite, total represents a maximum imaging time of a satellite on a single track, represents a start time of the kth task accepted by satellite j, represents an end time of the kth task accepted by satellite j, t trans represents a time of satellite attitude maneuvering.

[0093] Further, the task optimization module scores with the optimization objective as an evaluation function, including:

[0094] The evaluation function is used to evaluate a new round of task planning strategies, and the evaluation scores and strategy codes are stored in a program database together. The strategy code refers to a runnable Python program generated by the large model according to the prompt word. The program database stores these strategy codes and their corresponding evaluation scores. The prompt word is a string of text information designed by humans, describing the requirements of the task, as input to the large model.

[0095] Based on the island model and genetic algorithm, the program database is updated, and the prompt word is sampled and input to the large language model. The island model divides the program database into multiple independent sub-populations (islands), each island evolves programs independently, and periodically resets operations to maintain diversity and avoid local optimization. The genetic algorithm selects the best performing program within the island, and generates new program individuals through crossover and mutation. The new program is tested by the evaluator, and if it meets the requirements, it is stored in the program database. Excellent programs are used to generate new programs through sampling, promoting continuous optimization and improvement, and ultimately achieving the goal of finding the best solution.

[0096] The training set example is randomly selected from two orbit data of 63 Flock-4 remote sensing satellite constellations released by the North American Air Defense Command, 12 satellites are selected, a total of 5 times, a total of 5 groups of training sets, each group of training sets contains 12 satellites and 100 dynamically inserted stationary point targets. The simulation time is 90 minutes, about 1 orbit period, and the simulation start time is 1 Sep 2024 00:00:00. The satellite has only the ability to maneuver in the roll direction, the maximum roll angle is 32°, and a total of 150s is used for point target imaging per orbit period, and 10s is required to complete 1 point target imaging. The point targets are uniformly distributed between 45°N to 65°N and 150°E to 165°E. The target benefits are uniformly distributed between 1-5 and are all integers. The target function weight factor a = 0.8.

[0097] First, the earliest and first priority greedy algorithm of visible time is used to construct the program skeleton of distributed task planning, and then the large model is iteratively optimized for 500 rounds. The average value curve of the objective function on the 5 groups of training sets is as shown in Figure 2 Figure 2 It can be seen from that through 31 rounds of iteration, the objective function on the training set is improved from 0.666 to 0.668. After 278 rounds of iteration, the objective function is improved from 0.666 to 0.707, which is improved by 6.3% compared with the initial greedy algorithm.

[0098] The test set example is randomly selected from two orbit data of 63 Flock-4 remote sensing satellite constellations released by the North American Air Defense Command, 12, 18, 24, 30, and 36 satellites are selected, a total of 5 groups of training sets, each group of training sets contains 300 dynamically inserted stationary point targets, and the remaining parameter settings are consistent with the training set.

[0099] This embodiment uses the earliest and first priority greedy algorithm of visible time as the control group for experiment, takes the average value of the 5 groups of test sets, and compares the objective function G, the total benefit reward of the planned task in the task planning period and the average remaining available imaging time t remain of the satellite. The results of the two are compared as shown in Table 1.

[0100] Table 1 Comparison of results of greedy algorithm and large language model adaptive task planning algorithm

[0101]

[0102] As can be seen from the above table, the large language model self-adaptive task planning algorithm has obvious advantages in the target function, total task planning revenue and satellite average remaining available imaging time compared with the greedy algorithm, which increases by 8.4%, 9.2% and 5.3% respectively. Therefore, the remote sensing satellite group self-adaptive task planning method based on the large language model can be effectively extended to the scenario of large-scale satellites, improve the total revenue of task planning, and also reduce the consumption of total imaging time of satellites.

[0103] The embodiment has the following advantages:

[0104] Effectiveness. After simulation examples are verified, the embodiment can effectively perform multi-satellite distributed task planning on dynamic tasks, improve task revenue and reduce satellite imaging time consumption.

[0105] Timeliness. The traditional centralized task planning method has high algorithm complexity, and it is difficult to respond to the demand of dynamic tasks in real time due to the limitation of satellite-ground communication time. The distributed task planning method proposed in the embodiment has low time complexity and fast response speed for dynamic tasks.

[0106] Extensibility. After simulation examples are verified, the method proposed in the embodiment can be effectively extended to the scenario of large-scale satellites, has stronger generalization ability compared with the traditional method, and has better planning results.

[0107] The above-described embodiments only describe the preferred modes of the present application, and do not limit the scope of the present application. Without departing from the design spirit of the present application, various modifications and improvements to the technical solutions of the present application made by those skilled in the art shall fall within the protection scope determined by the claims of the present application.

Claims

1. A large language model-based remote sensing satellite swarm adaptive task planning method, characterized in that, The method comprises the following steps: Based on the decision variables, constraint conditions and optimization objectives of the remote sensing satellite group task planning, an optimization model is constructed; The optimization model is: wherein G represents an optimization objective function, represents a weight factor, represents a proportion of a total income of a planned task in a task planning period to a maximum income, represents a proportion of a total available imaging time of all satellites in a task planning period to a maximum imaging time, represents a satellite planning period, decision state, represents a total number of tasks, represents a total number of satellites, represents a minimum imaging time of a satellite, represents a maximum imaging time of a satellite, represents a start time of a first task accepted by a satellite, represents an end time of a first task accepted by a satellite, represents an end time of a first task accepted by a satellite, represents a satellite attitude maneuvering time;​​​ Based on the optimization model, a greedy algorithm is used to generate an initial task planning strategy; Through a large language model, the initial task planning strategy is iteratively optimized, a new round of task planning strategy is generated each time, and the optimization objective is used as an evaluation function to score, until the preset number of iterations is reached, the optimization is stopped, and the task planning strategy with the highest score is output; The optimization objective is used as an evaluation function to score, which includes: The evaluation function is used to evaluate the new round of task planning strategy, and the evaluation score and strategy code are stored in the program database together; wherein the strategy code refers to a runnable Python program generated by the large model according to the prompt word, and the prompt word is a string text information artificially designed to describe the requirements of the task as the input of the large model; Based on the island model and genetic algorithm, the program database is updated, and the prompt word is input into the large language model for sampling; wherein the island model divides the program database into multiple independent subpopulations, each subpopulation evolves programs independently, and periodically resets; the genetic algorithm selects the best performing program in the subpopulation, and generates new program individuals through crossover and mutation, and the new program is tested by the evaluator, and if it meets the requirements, it is stored in the program database; Based on the task planning strategy with the highest score, the remote sensing satellite group adaptive task planning is carried out.

2. The adaptive task planning method for remote sensing satellite groups based on a large language model according to claim 1, characterized in that: The decision variable is: satellite To the task Decision state; wherein the decision state includes: And accept the task; The constraint conditions include: the maximum imaging time of a single satellite track, the minimum imaging time of a single satellite, the imaging of a target by a satellite must be completed within the satellite's visible window, and the time interval between adjacent two imaging tasks must be greater than the satellite's attitude maneuvering time; The optimization objectives include: maximizing task revenue and maximizing satellite remaining available imaging time.

3. A large language model-based remote sensing satellite swarm adaptive task planning system, characterized in that, The system for implementing the adaptive task planning method for remote sensing satellite groups based on a large language model according to any one of claims 1-2, the system comprises a construction module, a strategy generation module, a task optimization module and an adaptive task planning module; The construction module is configured to construct an optimization model based on the decision variables, constraint conditions and optimization objectives of the remote sensing satellite group task planning; The optimization model is: wherein G represents an optimization objective function, represents a weight factor, represents a proportion of total revenue of a planned task in a task planning period to maximum revenue, represents a proportion of total available imaging time of all satellites in a task planning period to maximum imaging time, represents a satellite decision state, of a task, represents a total number of tasks, represents a total number of satellites, represents a minimum imaging time of a satellite, represents a maximum imaging time of a satellite, represents a start time of a first task accepted by a satellite, represents an end time of the first task accepted by the satellite, represents an end time of the first task accepted by the satellite, represents a time of a satellite attitude maneuver, represents a time of a satellite attitude maneuver. The strategy generation module is configured to generate an initial task planning strategy based on the optimization model using a greedy algorithm; The task optimization module is configured to iteratively optimize the initial task planning strategy through a large language model, generate a new round of task planning strategy each time, and use the optimization objective as an evaluation function to score, until the preset number of iterations is reached, the optimization is stopped, and the task planning strategy with the highest score is output; The task optimization module uses the optimization objective as an evaluation function to score, which includes: The evaluation function is used to evaluate a new round of task planning strategies, and the evaluation scores and strategy codes are stored in the program database together; wherein the strategy code refers to a runnable Python program generated by the large model according to a prompt word, and the prompt word is a string text information artificially designed to describe the requirements of the task as the input of the large model; The program database is updated based on an island model and a genetic algorithm, and the prompt word input is sampled and generated to the large language model; wherein the island model divides the program database into multiple independent sub-populations, each sub-population evolves the program independently, and is reset periodically; the genetic algorithm selects the best performing program within the sub-population, and generates new program individuals through crossover and mutation, and the new program is tested by the evaluator, and if it meets the requirements, it is stored in the program database; The adaptive task planning module is used for adaptive task planning of the remote sensing satellite group based on the task planning strategy with the highest score.

4. The adaptive task planning system of the remote sensing satellite group based on the large language model according to claim 3, characterized in that, The decision variable is: satellite To the decision state of the task ; wherein the decision state includes: And accept the task; The constraint conditions include: the maximum imaging time of a single satellite track, the minimum imaging time of a single satellite, the imaging of a satellite on a target must be completed within the satellite visible window, and the time interval between adjacent two imaging tasks needs to be greater than the satellite attitude maneuvering time; The optimization target includes: maximizing the task benefit and maximizing the remaining available imaging time of the satellite.

Images