Planning method and system for sky-ground measurement, operation and control tasks

Abstract

The invention relates to the technical field of task planning, and provides a planning method and system for a sky-ground-based measurement, operation and control task, and the method comprises the steps: constructing a sky-ground-based measurement, operation and control task planning environment model, which comprises the construction of a task set, an equipment set, a constraint condition set and an optimization target set; constructing a state space, an action space and a reward function; constructing a deep Q network model; training a depth Q network model based on the sky-ground measurement, operation and control task planning environment model, the state space, the action space and the reward function to obtain a task planning model; and inputting the acquired current state vector of the sky-foundation measurement, operation and control system into the task planning model to obtain a task allocation scheme and realize task planning. According to the method, the problem of space-foundation measurement, operation and control resource allocation under multiple constraints can be solved, local optimum is avoided, and the task completion rate and the equipment resource utilization rate are increased.

Description

Technical Field This disclosure belongs to the field of mission planning technology, and in particular relates to a planning method and system for space-ground measurement and control missions. Background Technology

[0001] In the field of aerospace engineering, the space-ground-based telemetry, tracking, and control system undertakes key tasks such as tracking, measuring, controlling, and transmitting data to spacecraft. Its mission planning needs to comprehensively consider multi-dimensional constraints, including the spatiotemporal coverage of telemetry and control resources, equipment performance parameters, mission priorities, energy limitations, and meteorological conditions. The rationality and efficiency of mission planning directly affect the operational safety of spacecraft and the quality of mission completion. Traditional methods for planning space-ground-based measurement, control, and operations (GC) missions mainly include rule-based planning methods and heuristic search methods. Rule-based planning methods rely on a pre-defined rule base, allocating and scheduling tasks based on their basic attributes. However, this method has significant limitations. When the number of tasks, equipment types, and constraints in GC scenarios increase, manually defined rules struggle to cover all situations, easily leading to rule conflicts or suboptimal planning results, and lacking adaptability to dynamically changing scenarios. Heuristic search methods, on the other hand, construct objective functions and search strategies to find optimal or near-optimal task planning schemes. While offering some improvement in planning performance compared to rule-based methods, these methods typically require complex mathematical modeling of the problem, and the search process is prone to getting trapped in local optima. Especially in space-ground-based GC missions, complex dependencies exist between tasks, and equipment resources face multi-dimensional constraints, significantly increasing the computational complexity of heuristic search methods, resulting in longer planning times and making it difficult to meet the real-time requirements of GC scenarios. With the surge in low-Earth orbit satellite constellations and commercial space missions, the diversity and complexity of space-based telemetry, tracking, and command (TT&C) missions have increased significantly. Traditional methods can no longer meet the planning requirements of "high dynamics, high concurrency, and multiple constraints" TT&C missions. Summary of the Invention

[0002] To address the aforementioned issues, this disclosure provides a planning method and system for space-ground measurement and control missions. The method employs an environment model to construct a state space, action space, and reward function to train a deep Q-network model for mission planning. This approach can handle resource allocation issues in space-ground measurement and control under multiple constraints, avoid getting trapped in local optima, and improve mission completion rate and equipment resource utilization.

[0003] The following is the technical content of this disclosure: A planning method for space-ground ground-based measurement and control missions, characterized by comprising: Construct a space-ground-based measurement, control, and operation mission planning environment model, including constructing a task set, equipment set, constraint set, and optimization objective set; Construct the state space, action space, and reward function; Construct a deep Q-network model; based on the aforementioned space-ground measurement and control mission planning environment model, state space, action space, and reward function, train the deep Q-network model to obtain the mission planning model; The current state vector of the acquired space-ground measurement and control system is input into the task planning model to obtain the task allocation scheme and realize task planning.

[0004] Furthermore, The tasks in the task set include the following attributes: Task types include: tracking and measurement tasks, data transmission tasks, and control command sending tasks; Task priority, which includes multiple levels of priority; Task start time window; Task duration; The set of resources required for the task.

[0005] Furthermore, The devices in the device set include the following attributes: Equipment number; Equipment types include: ground control stations, relay satellites, and data processing equipment; Equipment operating time window; The device can provide a collection of resources; Current device status; Equipment historical task execution records.

[0006] Furthermore, The conditions in the set of constraints include: The execution time of a task assigned to a device must fall entirely within the device's working time window; the set of resources required by the task must be included within the set of resources available to the device; and only one task can be executed by the same device within the same time period.

[0007] Furthermore, The state space is represented in vector form, and the state vector contains task-related features, device-related features, and environment-related features. The task-related features include: the number of tasks to be assigned, the remaining time window of each task to be assigned, the priority of each task to be assigned, and the scarcity of resources required by each task to be assigned. ; The equipment-related characteristics include the current status of each device, the remaining working time of each device, the total execution time of the tasks assigned to each device, and the type and quantity of resources occupied by each device; the environment-related characteristics include: the current time, the remaining time within the preset planning period, the historical task completion rate, and the historical device resource utilization rate.

[0008] Furthermore, The reward function consists of immediate reward, stage reward and terminal reward, and the total reward value is calculated by superimposing the current time type. The instant reward is obtained after the intelligent agent performs a valid action; the stage reward is calculated and obtained at the end of the preset stage time; and the terminal reward is calculated at the end of the preset planning period. The total reward value is:

[0009] in:

[0010]

[0011]

[0012] In the formula, For instant rewards; As a stage reward; For terminal rewards; These are the weighting coefficients; ; This represents the ideal value for equipment load balancing. and These are the weighting coefficients; These are the weighting coefficients; and ; The task completion rate for the entire cycle; The equipment resource utilization rate for the entire cycle; The total task execution cost for the entire cycle; The preset maximum total execution cost; The task priority weight completion rate for the entire cycle.

[0013] Furthermore, The training process of the deep Q-network model includes: Initialize the parameters of the main network and the target network of the deep Q-network model; Initialize the agent's current state vector. At each time step, select an action using the ε-greedy policy. After executing the action, update the task and device states, calculate the reward, generate the next state vector, and store the experience samples in the buffer. When the number of samples in the buffer reaches a preset threshold, samples are randomly sampled, the target Q value is calculated, and the main network parameters are updated using the mean squared error loss function. At preset update intervals, the main network parameters are copied to the target network; When the training objective is achieved, training is stopped and the main network parameters are saved.

[0014] Furthermore, After obtaining the trained task planning model, real-time task planning is achieved through the following steps: Collect real-time data, convert it into a state vector, and process it according to normalization; Input the real-time state vector into the task planning model, calculate the Q-value of all actions, filter actions that meet the hard constraints, select the action with the largest Q-value to generate the task assignment instruction, and execute an empty action if there are no actions that meet the constraints.

[0015] A planning system for space-ground measurement and control missions, characterized in that it includes: The environment model construction module is used to construct the planning environment model for space-ground-based measurement, control and operation missions, including constructing the task set, equipment set, constraint set, and optimization objective set; The training environment building module is used to build the state space, action space, and reward function; The model training module is used to construct a deep Q-network model; based on the aforementioned space-ground-based measurement and control mission planning environment model, state space, action space, and reward function, the deep Q-network model is trained to obtain the mission planning model; The task planning module is used to input the current state vector of the acquired space-ground measurement and control system into the task planning model to obtain the task allocation scheme and realize task planning.

[0016] Furthermore, Also includes: The model optimization module is used to calculate the task completion rate, equipment resource utilization rate, total task execution cost, and task priority weight completion rate within each planning cycle after the end of each planning cycle, based on the set of optimization objectives, in order to evaluate the current model performance. If the evaluation metrics do not meet the preset requirements, the deep Q-network model will be retrained.

[0017] Compared with the prior art, this disclosure has the following advantages: This disclosure uses a deep Q-network model for task planning. The state space integrates three core features—task, device, and environment—in a vector form. This achieves full-dimensional environmental perception to eliminate blind spots in planning information and avoids suboptimal decisions caused by missing information such as resources and historical data. It also solves the interference of features of different magnitudes on model training, ensures training stability, and allows the model to pay balanced attention to multi-dimensional factors, thereby improving the accuracy of task allocation schemes. Among them, the action space explicitly pre-screens effective actions with hard constraints such as "task execution time falls within the device's working window, the resources required by the task are included within the resources that the device can provide, and only one task is executed by the same device at the same time period". It also sets "empty actions" to deal with scenarios where there is no effective allocation, directly excluding invalid actions, greatly reducing the amount of computation for invalid decisions in the model, improving planning efficiency to meet real-time requirements, and avoiding resource mismatch by mapping action definitions to "task-device", thereby improving the effectiveness of single task allocation. The reward function consists of immediate rewards, stage rewards, and terminal rewards. Immediate rewards provide real-time feedback on the effectiveness of actions for rapid error correction. Stage rewards are based on task completion rate and equipment load balancing adjustment strategies within a stage. Terminal rewards comprehensively consider task completion rate, equipment resource utilization, total execution cost, and task priority weight completion rate within a cycle to ensure global optimization. Furthermore, the weights of each reward item can be flexibly adjusted. This achieves multi-layered optimization guidance from "real-time to stage to global" and can adapt to diverse scenarios such as "prioritizing urgent tasks" and "low-cost operation." It breaks through the limitations of traditional single optimization objectives and significantly improves the practicality and optimization of the planning scheme.

[0018] Other features and advantages of this disclosure will be set forth in the following description and will be apparent in part from the description or may be learned by practicing the disclosure. The objects and other advantages of this disclosure may be realized and obtained by means of the structures pointed out in the description, claims and drawings. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of this disclosure or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this disclosure. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 A schematic diagram of the method of the present invention is shown; Figure 2 This diagram illustrates the training and policy optimization process of a deep Q-network model. Figure 3 A schematic diagram of the overall process for automatic planning of space-ground measurement and control missions using a deep Q-network is shown. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of the embodiments of this disclosure clearer, the technical solutions of the embodiments of this disclosure will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this disclosure, and not all embodiments. Based on the embodiments of this disclosure, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this disclosure.

[0022] Figure 1 A schematic diagram of the method according to the present invention is shown. Specific implementation details of the present invention include: Step 1: Construct a space-ground-based measurement, control, and operation mission planning environment model; The elements involved in the planning of space-ground-based measurement, control, and operation missions are clearly defined, including the task set, equipment set, constraint set, and optimization objective set. A mission planning environment model is then constructed, as follows: (1) Definition of task set: The tasks of ground-based measurement and control are abstracted into a task set. Each of the tasks Includes task type and task priority Task start time window Task duration and the set of resources required for the task .

[0023] (2) Definition of equipment resource set: Let the equipment set be... ,in For the total number of devices, each device Includes the following attributes: Device ID, Device Type, and Device Operating Time Window. The equipment can provide a collection of resources. Current status of the equipment Equipment historical task execution records .

[0024] (3) Definition of constraint set: Constraints include hard constraints and soft constraints. Hard constraints are conditions that must be met, while soft constraints are conditions that should be met as much as possible. Hard constraints: ① Task Execution time It must fall completely on the equipment Working time window Inside, among which For the task Assigned to device The start time of execution after; ② Task Required resource set It must be fully included in the device. Available resource collection Within; ③ A single device can only execute one task at a time, that is, for a given device... If a task has already been assigned Its execution time is Then the newly assigned task Execution time and No overlap; Soft constraints: ① High-priority tasks are allocated resources and equipment first; ② Equipment load balancing, i.e., the difference in the total task execution time of each device is controlled within a preset threshold; ③ Minimize task execution cost, including equipment energy consumption cost, resource usage cost, etc. Let the task... By equipment The cost of execution is The total execution cost should be reduced as much as possible. ,in Indicates task Assigned to device , This indicates that the allocation is unassigned.

[0025] (4) Definition of Optimization Objectives: Optimization objectives include: ① Maximizing task completion rate, i.e., maximizing the proportion of successfully assigned and executed tasks to the total number of tasks; ② Maximizing equipment resource utilization, where equipment resource utilization = actual equipment working time / equipment working time window duration, and the overall equipment resource utilization is the average of all equipment utilization rates, which is then maximized; ③ Minimizing total task execution cost; ④ Maximizing task priority weight completion rate, i.e., the completion rate weight of high-priority tasks is higher than that of low-priority tasks, and the priority weight completion rate = ,in Indicates task Finish, This indicates that the task is incomplete.

[0026] Step 2: Construct the state space, action space, and reward function (1) State space construction: The state space is used to describe the current environmental state during the task planning process. It is represented in vector form, and the state vector is... , The state dimension specifically includes the following features: Task-related characteristics: number of tasks to be assigned, and remaining time window for each task to be assigned. - Current time, priority of each task to be assigned, and scarcity of resources required by each task to be assigned; ; Equipment-related characteristics: Current status of each piece of equipment, remaining working time of each piece of equipment. - Current time, total execution time of tasks assigned to each device, and type and quantity of resources occupied by each device; Environmental characteristics: current time, remaining time within the preset planning period, historical task completion rate, and historical equipment resource utilization rate; Normalize all features in the state vector and map the feature values ​​to... The interval is used to avoid the influence of features of different magnitudes on the algorithm.

[0027] (2) Action Space Construction: The action space is defined as the actions that the agent can execute, i.e., the actions assigned to the task. Let the action set be... ,in Indicates the task Assigned to device In addition, set up empty actions. This indicates that no task assignment will be performed at the current moment. The validity of an action must satisfy the hard constraints defined in step 1. When selecting an action, the set of valid actions that satisfy the hard constraints is first filtered out. If the set of valid actions is empty, an empty action is executed.

[0028] (3) Reward function construction: The reward function is used to evaluate the environmental feedback after the agent performs an action, guide the agent to learn the optimal policy, and the reward value is used to evaluate the environmental feedback after the agent performs an action. From instant rewards Stage rewards and terminal rewards The composition, and the specific calculation formula are as follows: Instant rewards When performing a non-empty action Afterwards, if the action satisfies the hard constraints and the task... Successfully assigned to the device ,but ,in and These are weighting coefficients, which can be adjusted according to actual needs; if the action does not meet the hard constraints, then... , This is the penalty coefficient; if an empty action is performed, then... , A small penalty coefficient is used to avoid the agent frequently performing empty actions; Stage Rewards Calculated at the end of each preset phase period, the phase reward is based on the task completion rate within that phase. Equipment resource utilization rate and equipment load balancing Related, the calculation formula is as follows ,in These are the weighting coefficients, and , This represents the ideal value for equipment load balancing. Equipment load balancing =1 - Standard deviation of total task duration assigned to each device / Average of total task duration assigned to each device; Terminal rewards The terminal reward is calculated at the end of the entire planning period, taking into account the achievement of optimization goals throughout the period. The calculation formula is as follows: ,in These are the weighting coefficients, and ; The task completion rate for the entire cycle. For the entire cycle of equipment resource utilization, The total task execution cost for the entire cycle. The preset maximum total execution cost, The task priority weight completion rate for the entire cycle; Total reward value At non-stage and non-terminal moments, , At the stage moment, At the terminal moment, all three rewards are calculated.

[0029] Step 3: Construct a deep Q-network model The deep Q-network model employs a structure combining convolutional neural networks and fully connected neural networks to fit the action value function. , indicating the state Next action The expected cumulative reward structure is as follows: (1) Input layer: receives the normalized state vector The input dimension is the dimension of the state space. .

[0030] (2) Convolutional Layers: Set 2-3 convolutional layers to extract local features from the state vector. The first convolutional layer uses 32 3×3 convolutional kernels with a stride of 1 and ReLU activation function; the second convolutional layer uses 64 3×3 convolutional kernels with a stride of 1 and ReLU activation function; if a third convolutional layer is set, 128 3×3 convolutional kernels are used with a stride of 1 and ReLU activation function. Each convolutional layer is followed by a max pooling layer with a kernel size of 2×2 and a stride of 2 to reduce the number of parameters and prevent overfitting.

[0031] (3) Fully connected layers: Set 2-3 fully connected layers to map the features extracted by the convolutional layers to the action value space. The first fully connected layer has 256 neurons and uses ReLU as the activation function; the second fully connected layer has 128 neurons and uses ReLU as the activation function; if a third fully connected layer is set, the number of neurons is 64 and ReLU is used as the activation function.

[0032] (4) Output layer: The number of neurons in the output layer is equal to the size of the action space. Each neuron corresponds to the Q-value of an action, and a linear activation function is used.

[0033] Step 4: Model Training and Policy Optimization The deep Q-network is trained using experience replay and target network techniques. The specific steps are as follows: (1) Step 4.1: Initialize parameters: Initialize the master network parameters of the deep Q network and target network parameters Initially The network weights are initialized using a random normal distribution, and the biases are initialized using a constant. Initialize the experience replay buffer with a capacity of N to store experience samples generated by the agent's interactions with the environment. ,in for The state at any given moment, for Actions performed at all times for Rewards earned at any time for The state at any given moment, As a termination marker, This indicates the end of the current planning cycle. This indicates that the process is not yet complete. Set training parameters: learning rate Discount Factor Target network update cycle Training batch size Number of training iterations .

[0034] (2) Step 4.2: Intelligent agent interacts with environment to collect experience Initialize the current state Based on the environment model constructed in step 1, the initial task set, device set state, and environment state are obtained, and an initial state vector is generated. And normalize; For each training iteration cycle, the time step ,T The maximum time step per iteration cycle; use - Greedy strategy for selecting actions: generating random numbers , The exploration rate is initially set to 1.0, and decreases linearly to 0.01 with each training iteration. Then, randomly select an action from the set of valid actions. ;like The current state is then calculated through the main network. Given the Q values ​​of each valid action, select the action with the largest Q value. ; Execute action Update the task status and device status based on the environment model from step 1, and calculate the immediate reward. If the current time step is a stage time node, calculate the stage reward. If the current time step is the end of the cycle, calculate the terminal reward. and set ,otherwise Total Rewards ; Get the next state Generate the next state vector based on the updated task, device, and environment states. And normalize; empirical samples Store the sample in the experience replay buffer. If the buffer is full, replace the earliest stored sample according to the first-in-first-out principle.

[0035] (3) Step 4.3: Network parameter update When the number of samples in the experience replay buffer reaches a preset threshold, network parameter updates begin. Random sampling from the experience playback buffer A sample of experiences ; Calculate the target Q value For each sampled sample, if Then the target Q value ;like Then the next state is calculated through the target network. Take the maximum Q value from all valid actions. Target Q value ; Calculate the predicted Q-value of the main network. That is, the main network is in state Next action Q-value output; Define the loss function: Use the mean squared error loss function, and the loss value is: ; Optimize network parameters: Employ stochastic gradient descent or adaptive moment estimation optimizers, based on the loss value. Calculate the gradient and update the main network parameters along the gradient descent direction. Minimize the loss function.

[0036] (4) Step 4.4: Target network parameter synchronization Every time Each time step will display the current parameters of the main network. Copy to the target network and update the target network parameters. This ensures the stability of the target Q value and avoids oscillations during training.

[0037] (5) Step 4.5: Training termination judgment Repeat steps 4.2 through 4.4 until completion. Training is stopped when the average loss value of multiple consecutive iterations is less than a preset threshold and the optimization target indicators such as task completion rate and device resource utilization on the test set are stable within a preset range. The final main network parameters are then saved as the trained deep Q-network model.

[0038] The training and policy optimization process for the deep Q-network model is attached. Figure 2 .

[0039] Step 5: Real-time task planning and dynamic adjustment (1) Step 5.1: Real-time status acquisition and preprocessing In the practical application phase, real-time acquisition of mission data, equipment data, and environmental data from the space-ground-based measurement, control, and operation system is conducted. According to the definition of state space in step 2, the collected real-time data is converted into a state vector, and all features in the vector are normalized to obtain the current real-time state. .

[0040] (2) Step 5.2: Action selection and task allocation Real-time status Input the trained deep Q-network master model and calculate the Q-value for all possible actions; Filter out valid actions that meet the hard constraints in step 1. If a valid action exists, select the action with the largest Q value as the optimal task allocation scheme and generate a task allocation instruction. If no valid action exists, execute an empty action and wait for the next planning time step to re-collect the state and select actions.

[0041] (3) Step 5.3: Execute monitoring and dynamic adjustment Send task assignment instructions to the corresponding ground-based and satellite-based devices, and monitor the task execution status and completion status of the devices; In case of an emergency, immediately re-collect the current status and repeat steps 5.1 to 5.2 to quickly generate a new task allocation plan, realize dynamic adjustment of task planning, ensure that the system can adapt to emergency scenarios, and guarantee the smooth execution of tasks.

[0042] (4) Step 5.4: Evaluation of planning effect and model iteration After each planning cycle ends, based on the set of optimization objectives in step 1, calculate the task completion rate, equipment resource utilization rate, total task execution cost, task priority weight completion rate, and other indicators for that cycle, and evaluate the effectiveness of the current planning method. If the evaluation indicators do not meet the preset requirements, or if the system's task type, number of devices, or constraints change significantly, then the experience samples within that period are collected and added to the experience playback buffer. The deep Q-network model is then incrementally trained according to the methods in steps 4.3 to 4.5 to achieve iterative optimization of the model and improve its adaptability to new scenarios.

[0043] The overall process of automatic planning for space-ground measurement and control missions based on deep Q-networks is attached. Figure 3 .

[0044] Based on the method of the present invention, this disclosure also provides a system corresponding to the above method, which includes: The environment model construction module is used to construct the planning environment model for space-ground-based measurement, control and operation missions, including constructing the task set, equipment set, constraint set, and optimization objective set; The training environment building module is used to build the state space, action space, and reward function; The model training module is used to construct a deep Q-network model; based on the aforementioned space-ground-based measurement and control mission planning environment model, state space, action space, and reward function, the deep Q-network model is trained to obtain the mission planning model; The task planning module is used to input the current state vector of the acquired space-ground measurement and control system into the task planning model to obtain the task allocation scheme and realize task planning.

[0045] Although the present disclosure has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present disclosure.

Claims

1. A method for planning a space-ground based measurement control mission, characterized in that, The application relates to a method for constructing a task planning model for a space-ground measurement and control system. The method comprises the following steps: constructing a space-ground measurement and control task planning environment model, including constructing a task set, a device set, a constraint condition set and an optimization target set; constructing a state space, an action space and a reward function; constructing a deep Q network model; training the deep Q network model based on the space-ground measurement and control task planning environment model, the state space, the action space and the reward function to obtain a task planning model; 2. The method according to claim 1, wherein, inputting a current state vector of the space-ground measurement and control system into the task planning model to obtain a task allocation scheme and realize task planning. The tasks in the task set comprise the following attributes: a task type, including a tracking measurement task, a data transmission task and a control instruction sending task; a task priority, including multiple levels of priority; a task start time window; a task required time length; 3. The method of claim 1, wherein, a task required resource set. The devices in the device set comprise the following attributes: a device number; a device type, including a ground measurement and control station, a relay satellite and a data processing device; a device working time window; a device available resource set; a device current state; 4. The method according to claim 1, wherein, a device historical task execution record. The constraint conditions in the constraint condition set comprise the following:

5. The method according to claim 1, wherein, after a task is allocated to a device, the execution time of the task must completely fall within the working time window of the device, the task required resource set must be contained in the available resource set of the device, and the same device can only execute one task in the same time period. ； The state space is represented in a vector form, and the state vector comprises task related features, device related features and environment related features; the task related features comprise the number of tasks to be allocated, the remaining time window of each task to be allocated, the priority of each task to be allocated and the scarcity degree of the resource required by each task to be allocated; 6. The method according to claim 1, wherein, the device related features comprise the current state of each device, the remaining working time of each device, the total execution time of the allocated tasks of each device and the type and quantity of the occupied resources of each device; and the environment related features comprise the current time, the remaining time in a preset planning period, a historical task completion rate and a historical device resource utilization rate. The reward function is composed of an instant reward, a stage reward and a terminal reward, and a total reward value is calculated according to the current time type; the instant reward is obtained after an intelligent agent executes an effective action, the stage reward is calculated when a preset stage time ends, and the terminal reward is calculated when a preset planning period ends; the total reward value is as follows: In the formula, is an instant reward; is a stage reward; is a terminal reward; is a weight coefficient; ; is an ideal value of device load balancing degree; and is a weight coefficient; is a weight coefficient; and ; is a task completion rate of the entire period; is a device resource utilization rate of the entire period; is a total task execution cost of the entire period; is a preset maximum total execution cost; is a task priority weight completion rate of the entire period.

7. The method of claim 1, wherein, wherein: the training process of the deep Q network model comprises the following steps: initializing the parameters of a main network and a target network of the deep Q network model; initializing a current state vector of the intelligent agent, selecting an action at each time step through an epsilon-greedy strategy, updating the states of the tasks and the devices after the action is executed, calculating a reward, generating a next state vector, and storing experience samples in a buffer; when the number of samples in the buffer reaches a preset threshold, randomly sampling the samples, calculating target Q values, and updating the parameters of the main network by using a mean square error loss function; every interval of a preset update period, the parameters of the main network are copied to the target network; 8. The method according to claim 1, wherein, when a training target is achieved, the training is stopped and the parameters of the main network are saved. After the trained task planning model is obtained, real-time task planning is realized through the following steps: Collect real-time data, convert into state vector and normalize; Input real-time state vector into task planning model, calculate all action Q values, filter actions meeting hard constraints, select action with maximum Q value to generate task assignment instruction, if no action meeting constraints, execute null action.

9. A system for planning a mission of a space-ground base measurement control task, characterized by, Comprise: An environment model construction module for constructing a space-ground measurement operation control task planning environment model, including constructing a task set, a device set, a constraint condition set and an optimization target set; A training environment construction module for constructing a state space, an action space and a reward function; A model training module for constructing a deep Q network model; Train the deep Q network model based on the space-ground measurement operation control task planning environment model, the state space, the action space and the reward function to obtain a task planning model; A task planning module for inputting a collected current state vector of a space-ground measurement operation control system into the task planning model to obtain a task assignment scheme and realize task planning.

10. The system of claim 9, wherein, Further comprise: A model optimization module for calculating a task completion rate, a device resource utilization rate, a total task execution cost and a task priority weight completion rate in each planning period according to the optimization target set to evaluate a current model effect after the end of each planning period; If evaluation indexes do not meet preset requirements, retrain the deep Q network model.

Images