A method and system for scheduling computing resources for space science satellite ground data processing jobs

Abstract

This application provides a computational resource scheduling method and system for ground data processing operations of space science satellites. The method includes: modeling the scheduling problem as a Markov decision process; using a near-end policy optimization algorithm as the core algorithm of a deep reinforcement learning model, taking the augmentation state as input and outputting scheduling actions; executing scheduling according to the scheduling actions; the augmentation state includes predicted data and the current real-time system state; the predicted data includes the predicted downlink data volume and task execution duration, the corresponding CPU core requirements, and memory resource requirements; before executing scheduling, a resource adaptation check is performed; if resources are sufficient, they are directly allocated; if resources are insufficient and the task is of high priority, a preemption mechanism is triggered to preempt resources from lower-priority regular tasks; if there are no preemption conditions, the action is invalid. The advantages of this application are: effectively guaranteeing high-timeliness tasks and fully meeting the requirements of space science missions for low latency and timely processing.

Description

Technical Field

[0001] This application belongs to the field of ground data processing and resource management for space science satellites, and specifically relates to a method and system for scheduling computing resources for ground data processing operations of space science satellites. Background Technology

[0002] With the continuous advancement of space science satellite technology, space science satellites generate a large amount of observational data during their on-orbit operation, placing higher demands on the processing capabilities and resource scheduling efficiency of ground data systems. Different space science satellites exhibit significant differences in observation modes, payload configurations, and data product types, requiring ground data processing systems to simultaneously handle heterogeneous tasks of multiple types and objectives. Under conditions of limited computing resources and a complex system operating environment, how to efficiently organize and rationally schedule satellite data processing tasks is a critical challenge. In actual operation, space science satellite ground data processing tasks exhibit significant dynamic, heterogeneous, and highly constrained characteristics. On the one hand, data processing tasks arrive continuously and dynamically as satellite data is downlinked, with significant differences in computing resource requirements and execution durations among different tasks, demonstrating obvious heterogeneity. On the other hand, the available computing resources of the ground system also change continuously during task execution. Furthermore, missions such as astronomical satellites involve opportunistic observation targets, and their data processing tasks have strict timeliness requirements, needing to complete processing and analysis within a very short time to ensure the value of scientific results. For example, in the Einstein Probe (EP) mission, astronomical alert data transmitted via VHF and BeiDou communication links typically requires processing and result generation within a very short time after reception to support multi-band collaborative observation needs. These stringent requirements for low-latency processing capabilities pose a greater challenge to the resource scheduling capabilities of ground data processing systems. Existing scheduling methods generally suffer from insufficient adaptability to dynamic changes, difficulty in simultaneously optimizing multiple objectives, and limited ability to guarantee high-priority tasks when facing such complex scenarios. Therefore, there is an urgent need for a computational resource scheduling method that can adapt to dynamic environments, possess efficient resource allocation capabilities, and meet the demands of time-sensitive tasks. Summary of the Invention

[0003] The purpose of this application is to overcome the shortcomings of existing technologies, such as insufficient adaptability to dynamic changes, difficulty in simultaneously optimizing multiple objectives, and limited ability to guarantee high-priority tasks.

[0004] To achieve the above objectives, this application proposes a computational resource scheduling method for ground data processing operations of space science satellites, comprising: The scheduling problem is modeled as a Markov decision process; The proximal policy optimization algorithm is used as the core algorithm of the deep reinforcement learning model. The input is the reinforcement state and the output is the scheduling action. Execute scheduling based on scheduling actions; The enhanced status includes predicted data and the current real-time status of the system; the predicted data includes the predicted downlink data volume and task execution duration, the corresponding CPU core requirements and memory resource requirements; Before the execution scheduling, a resource adaptation check is performed. If resources are sufficient, they are allocated directly. If resources are insufficient and it is a high-priority task, a preemption mechanism is triggered to preempt resources from low-priority regular tasks. If there are no preemption conditions, the action is invalid.

[0005] As an improvement to the above method, the initial values ​​of the parameters of the deep reinforcement learning model are obtained by pre-training the reinforcement learning policy network using an expert scheduling strategy, including: Shortest job priority is selected as the expert strategy. Full scheduling is performed on the simulation task set, and the state-action trajectory of each step is recorded to form an expert trajectory set. A convolutional neural network is used as the imitation learning model, with the system state as input and expert actions as supervision labels. The network is trained using mini-batch stochastic gradient descent and cross-entropy loss function.

[0006] As an improvement to the above method, the process of obtaining the prediction data includes: Input the historical task time-series feature vectors into the trained LSTM network to output predicted data; The historical task time-series feature vector includes time embedding features, historical downlink data volume, execution duration, payload type, and task priority.

[0007] As an improvement to the above method, the current real-time state of the system is as follows: The system state is encoded into a multi-channel two-dimensional matrix, which is divided into three parts: The cluster resource allocation area is used to represent the resource usage and task characteristics at each time step; The waiting queue area is used to represent the resource requirements, duration, and priority of tasks to be scheduled; The backlog queue area represents the number of tasks waiting to be scheduled that exceeds the capacity of the waiting queue.

[0008] As an improvement to the above method, the reward function of the deep reinforcement learning model is: ; Where λ is the resource utilization rate weight; κ is the priority intensity coefficient; For the task Priority; For the task The time already spent in the area, For the task arrival time, For a moment System time; U(t) This represents the average resource utilization rate of the cluster. For a moment The set of all incomplete tasks in the system.

[0009] As an improvement to the above method, during the operation of the deep reinforcement learning model, the policy network and value network are iteratively updated, including: Collect trajectory data; Calculate the cumulative discount reward and generalized advantage estimate, and update the value network; The policy network is updated based on the pruning objective function, and the interaction, sampling and updating are performed cyclically until the policy converges.

[0010] This application provides a computational resource scheduling system for ground data processing operations of space science satellites, implemented based on the above method. The system includes: The scheduling module is used to take the augmentation state as input and output the scheduling action, with the proximal policy optimization algorithm as the core algorithm of the deep reinforcement learning model. The execution scheduling module is used to perform scheduling according to the scheduling action; before the execution scheduling, a resource adaptation check is performed. If resources are sufficient, they are allocated directly; if resources are insufficient and it is a high-priority task, a preemption mechanism is triggered to preempt low-priority regular task resources. If there is no preemption condition, the action is invalid. An augmented state module is constructed to build the augmented states required for the input of the deep reinforcement learning model. The augmented states include the predicted data and the current real-time state of the system. The predicted data includes the predicted downlink data volume and task execution duration, the corresponding CPU core requirements, and memory resource requirements.

[0011] Compared with existing technologies, the advantages of this application are: 1. Effectively guarantee high-timeliness tasks: This method can ensure that the on-time completion rate of high-priority opportunistic tasks (such as astronomical eruption warnings) reaches 100%, fully meeting the requirements of space science tasks for low latency and timely processing.

[0012] 2. Improve system resource utilization: Through forward-looking resource reservation and efficient scheduling decisions, this invention can allocate and utilize limited computing resources more rationally, avoiding resource idleness or excessive competition, thereby improving the overall operational efficiency of the ground system.

[0013] 3. This invention solves the problem of large oscillations in the early stages of deep reinforcement learning. By combining high-quality initialization from imitation learning with forward-looking supplementation from the prediction module, the model obtains high initial rewards and low turnaround time in the early stages of iteration, resulting in smooth curves and rapid convergence. Attached Figure Description

[0014] Figure 1 The diagram shows a flowchart of a computational resource scheduling method for ground data processing operations of space science satellites. Figure 2 The diagram shown is a flowchart of the computational resource demand forecasting process. Figure 3 The figure shows the state-space modeling representation of the scheduling environment; Figure 4 The diagram shows the interaction logic of deep reinforcement learning. Detailed Implementation

[0015] The technical solution of this application will be described in detail below with reference to the accompanying drawings.

[0016] In the process of ground data processing for space science satellites, satellite data processing tasks arrive continuously and dynamically during the downlink data process. Different tasks exhibit significant differences in resource requirements, execution time, and priority, necessitating efficient scheduling under limited computing resources. This application constructs an intelligent scheduling method that integrates imitation learning, computing resource requirement prediction, and deep reinforcement learning to model dynamically arriving satellite data processing tasks as Markov decision processes (MDPs).

[0017] Example 1 This application provides a computational resource scheduling method for ground data processing operations of space science satellites, including: Step 1: Satellite Mission Modeling and State Encoding: Data processing tasks are divided into two categories: one is routine planned tasks triggered by data transmission and telemetry channels; the other is high-priority opportunistic burst tasks triggered by BeiDou or VHF links. Core attributes such as task request arrival time, processing deadline, execution duration, CPU / memory resource requirements, priority, and task type are extracted. The real-time task status is encoded into a multi-channel two-dimensional matrix representation, including a cluster resource allocation area, a waiting queue area, and a backlog queue area. This comprehensively depicts node resource occupancy, characteristics of tasks awaiting scheduling, and queue backlog, forming a standardized state input that can be efficiently processed by deep networks.

[0018] Step 2: Calculate resource demand forecast: Based on historical task execution data, construct a time series forecast model to predict the task load and resource demand within the future time window; the forecast includes the task data volume, execution time and resource demand change trend, and the forecast results are mapped to future resource occupancy; by introducing the forecast results as additional information into the scheduling decision process, the scheduling strategy becomes forward-looking and reduces the scheduling performance fluctuation caused by task suddenness.

[0019] Step 3: Imitation Learning Strategy Initialization: By constructing an expert scheduling strategy based on heuristic rules, scheduling simulation is performed on historical task data to generate state-action decision sample data; supervised learning methods are used to train the policy network so that the model can learn the scheduling behavior under the expert strategy, thereby obtaining the initial policy model; this step is used to narrow the search space of reinforcement learning and improve the efficiency and stability of subsequent training.

[0020] Step 4: Reinforcement Learning Scheduling Strategy Optimization: The satellite data processing task scheduling process is modeled as a Markov decision process, defining the system state, scheduling actions, and reward function. The scheduling actions include selecting tasks to be executed from the task queue or performing no-operation. The reward function comprehensively considers factors such as task turnaround time, task priority, and resource utilization. A policy gradient-based reinforcement learning algorithm is used to train and optimize the scheduling strategy, enabling the system to continuously learn the optimal scheduling strategy in a dynamic environment.

[0021] Step 5: Scheduling Result Update and Feedback: In actual operation, the scheduler receives newly arriving tasks at each time step and makes scheduling decisions based on the current augmented state using the trained model. For high-priority opportunistic tasks, if the system does not have sufficient idle resources, it can preempt currently executing regular tasks according to preset rules to ensure timely processing. Preempted regular tasks will be re-added to the scheduling queue for subsequent reallocation by the scheduler.

[0022] Example 2 like Figure 1 The diagram shown is a flowchart of the scheduling method of this invention. The computational resource scheduling method for ground data processing operations of space science satellites provided in this application addresses the computational resource scheduling scenario of ground data processing operations for space science satellites. It proposes an integrated online scheduling method that combines imitation learning, computational resource demand prediction, and deep reinforcement learning, taking into account the characteristics of dynamic task arrival, strong heterogeneity, strict time constraints, limited resources, and significant load fluctuations. The satellite data processing tasks processed in this application include regular planned tasks and high-priority opportunistic burst tasks, with resources covering computing resources such as CPU and memory. The scheduling process must meet constraints such as resource capacity, task timeliness, and priority preemption. Detailed implementation steps are as follows: Step 1: Satellite Data Processing Task and System Resource Modeling Formal definitions are provided for various data processing tasks triggered by downlink from space science satellites and ground cluster computing resources, and a basic mathematical model for scheduling problems is constructed.

[0023] Define task set Each task includes attributes such as arrival time, deadline, execution duration, resource requirements, priority, and task type.

[0024] Define cluster resource set Each node contains the total resource capacity and the current resource usage.

[0025] The scheduling problem is modeled as a Markov Decision Process (MDP), represented by a quintuple M =<S,A,P,R,γ> This provides a foundation for subsequent reinforcement learning decisions.

[0026] Step 2: Initialize the training of the imitation learning module's execution strategy. Using an expert scheduling strategy to pre-train the reinforcement learning policy network can quickly obtain an initial policy that is close to that of an expert, thereby improving convergence speed and training stability.

[0027] Shortest Job First (SJF) was selected as the expert strategy. A complete scheduling process was performed on the simulation task set, recording the state-action trajectory at each step to form an expert trajectory set. .

[0028] A convolutional neural network is used as the imitation learning model, with the system state s as input and the expert action a as the supervision label.

[0029] The network is trained using mini-batch stochastic gradient descent (SGD) and cross-entropy loss function.

[0030] Calculate strategy similarity ,when Training is stopped and parameters are saved when the value is ≥0.85. Used as initialization parameters for the deep reinforcement learning module.

[0031] Step 3: Calculate resource requirements and perform time series forecasting. As attached Figure 2 As shown, LSTM networks are used to predict future task loads and resource requirements, providing forward-looking information for scheduling decisions and mitigating short-sighted decision-making.

[0032] Construct a historical task time-series feature vector, which includes time embedding feature TimeEmb, historical downlink data volume Di, execution duration Ti, payload type, and task priority Pi.

[0033] Input the feature sequences from the past k time windows into the LSTM network, and output the downlink data volume for the next time step. With execution time A weighted multi-task loss function is used to balance the training of the two objectives, avoiding optimization bias caused by differences in scales.

[0034] The prediction results are transformed into CPU and memory requirements through mapping, and the predicted resource requirement vector is output. .

[0035] in, and These represent the predicted downlink data volume (in GB) and the task execution time (in seconds), respectively. and These represent the corresponding CPU core requirements and memory resource requirements, respectively. The above prediction results, as prior information, together with the current system state, constitute the enhanced state representation, as detailed in the appendix. Figure 3 .

[0036] Step 4: Constructing the augmented state space for the deep reinforcement learning model By integrating the current real-time system state with the output of step 3, an enhanced state representation containing forward-looking information is constructed. (See attached...) Figure 3 As shown, the system state is encoded into a multi-channel two-dimensional matrix, which is divided into three parts: Cluster resource allocation area: Displays resource usage and task characteristics at each time step; Waiting queue area: Displays the resource requirements, duration, and priority of tasks awaiting scheduling; Backlog queue area: Represents the number of tasks waiting to be scheduled that exceeds the capacity of the waiting queue.

[0037] The future resource demand forecast output from step 3 is then incorporated into the current state to form an enhanced state. , as input to the policy network.

[0038] Step 5: Deep reinforcement learning model performs action selection and resource scheduling A policy optimization scheme is designed based on the Proximal Policy Optimization (PPO) algorithm. Under enhanced conditions, it outputs scheduling actions to complete task allocation and resource preemption. (Appendix) Figure 4 The interaction logic for deep reinforcement learning.

[0039] Action space is defined as , Indicates an empty action. This indicates the k-th task in the scheduling waiting queue.

[0040] Perform resource compatibility checks before executing actions: If resources are sufficient, they will be allocated directly; if resources are insufficient and the task is of high priority, a preemption mechanism will be triggered to preempt resources from lower-priority regular tasks; if there are no preemption conditions, the action will be invalid and time will proceed.

[0041] It supports multi-task scheduling in a single time step, allowing the agent to continuously execute valid actions until an empty or invalid action is performed.

[0042] Step 6: Reward Function Calculation and Strategy Evaluation By taking into account task delay, priority, and resource utilization, immediate rewards are calculated to guide strategy optimization.

[0043] Calculate the negative reward for task retention and penalize tasks that are not completed for an extended period. Introduce a priority weighting coefficient. Strengthen the protection of high-priority tasks. Calculate the average resource utilization of the cluster. U(t) The final reward function is obtained as follows:

[0044] Where λ is the resource utilization rate weight, and κ is the priority intensity coefficient. For the task priority ( (The larger the value, the higher the priority). For a moment The set of all incomplete tasks in the system (including running, waiting queue, and backlog tasks). For the task The time spent in the area ( For the task arrival time, For a moment The penalty (system time) accumulates with the task dwell time, prompting the agent to shorten the task turnaround time.

[0045] Step 7: Iterative Update of Policy Network and Value Network The PPO algorithm is optimized using a near-end strategy to stabilize the update strategy and maximize long-term returns.

[0046] The intelligent agent interacts with the environment and collects trajectory data. .

[0047] Calculate the discounted cumulative reward and the generalized advantage estimate (GAE), update the value network, and minimize the difference between the predicted value and the actual reward.

[0048] The policy network is updated based on a pruning objective function, constraining the policy update magnitude to avoid gradient explosion and training oscillations. The process of interaction, sampling, and updating is repeated until the policy converges.

[0049] Example 3 This application also provides a computational resource scheduling system for ground data processing operations of space science satellites, implemented based on the method described above. The system includes: The scheduling module is used to take the augmentation state as input and output the scheduling action, with the proximal policy optimization algorithm as the core algorithm of the deep reinforcement learning model. The execution scheduling module is used to perform scheduling according to the scheduling action; before the execution scheduling, a resource adaptation check is performed. If resources are sufficient, they are allocated directly; if resources are insufficient and it is a high-priority task, a preemption mechanism is triggered to preempt low-priority regular task resources. If there is no preemption condition, the action is invalid. An augmented state module is constructed to build the augmented states required for the input of the deep reinforcement learning model. The augmented states include the predicted data and the current real-time state of the system. The predicted data includes the predicted downlink data volume and task execution duration, the corresponding CPU core requirements, and memory resource requirements.

[0050] The technical innovations of this invention include: Forward-looking scheduling decision mechanism: By constructing an enhanced state representation that includes future load information, it effectively solves the short-sighted decision problem caused by traditional reinforcement learning schedulers that rely only on the current state.

[0051] Efficient and stable policy learning: An imitation learning pre-training mechanism was designed to initialize the deep reinforcement learning policy network using the prior knowledge of expert policies (such as SJF), which accelerates the convergence speed of the model, reduces performance fluctuations in the early stage of training, and improves the engineering usability and stability of the algorithm.

[0052] Satellite mission image-based status representation mechanism: The resource occupation, waiting queue, backlog of tasks, priority and timeliness constraints of satellite data processing tasks are uniformly encoded into multi-channel two-dimensional image status, realizing the structured expression of high-dimensional information of scheduling scenarios.

[0053] This application may also provide a computer device, including: at least one processor, memory, at least one network interface, and a user interface. The various components in this device are coupled together via a bus system. It is understood that the bus system is used to implement communication between these components. In addition to a data bus, the bus system also includes a power bus, a control bus, and a status signal bus.

[0054] The user interface can include a display, keyboard, or clicking device. Examples include a mouse, trackball, touchpad, or touchscreen.

[0055] It is understood that the memory in the embodiments disclosed in this application may be volatile memory or non-volatile memory, or may include both volatile and non-volatile memory. The non-volatile memory may be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), or flash memory. The volatile memory may be random access memory (RAM), which is used as an external cache. By way of example, but not limitation, many forms of RAM are available, such as Static Random Access Memory (SRAM), Dynamic Random Access Memory (DRAM), Synchronous DRAM (SDRAM), Double Data Rate SDRAM (DDRSDRAM), Enhanced Synchronous DRAM (ESDRAM), Synchlink DRAM (SLDRAM), and Direct Rambus RAM (DRRAM). The memories described herein are intended to include, but are not limited to, these and any other suitable types of memory.

[0056] In some implementations, the memory stores elements such as executable modules or data structures, or subsets thereof, or extended sets thereof: operating systems and applications.

[0057] The operating system includes various system programs, such as the framework layer, core library layer, and driver layer, used to implement various basic business functions and handle hardware-based tasks. The application programs include various applications, such as media players and browsers, used to implement various application functions. Programs implementing the methods of the embodiments of this disclosure can be included in the application programs.

[0058] In the above embodiments, the processor can also invoke programs or instructions stored in memory, specifically programs or instructions stored in an application program, for the following purposes: Follow the steps described above.

[0059] The above methods can be applied to or implemented by a processor. The processor may be an integrated circuit chip with signal processing capabilities. During implementation, each step of the above methods can be completed by integrated logic circuits in the processor's hardware or by software instructions. The processor can be a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. It can implement or execute the methods, steps, and logic diagrams disclosed above. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the disclosed methods can be directly implemented by a hardware decoding processor, or by a combination of hardware and software modules in the decoding processor. The software modules can reside in random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, registers, or other mature storage media in the art. This storage medium is located in memory; the processor reads information from the memory and, in conjunction with its hardware, completes the steps of the above methods.

[0060] It is understood that the embodiments described in this application can be implemented using hardware, software, firmware, middleware, microcode, or a combination thereof. For hardware implementation, the processing unit can be implemented in one or more application-specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field-programmable gate arrays (FPGAs), general-purpose processors, controllers, microcontrollers, microprocessors, other electronic units for performing the functions described in this application, or combinations thereof.

[0061] For software implementation, the technology of this application can be implemented by executing the functional modules (e.g., procedures, functions, etc.) of this application. The software code can be stored in memory and executed by a processor. The memory can be implemented in the processor or outside the processor.

[0062] This application may also provide a non-volatile storage medium for storing a computer program. When the computer program is executed by a processor, it can implement the steps in the above method embodiments.

[0063] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application and are not intended to limit it. Although this application has been described in detail with reference to the embodiments, those skilled in the art should understand that modifications or equivalent substitutions to the technical solutions of this application do not depart from the spirit and scope of the technical solutions of this application, and should all be covered within the scope of the claims of this application.

Claims

1. A method for scheduling computational resources for ground data processing operations of space science satellites, comprising: The scheduling problem is modeled as a Markov decision process; The proximal policy optimization algorithm is used as the core algorithm of the deep reinforcement learning model. The input is the reinforcement state and the output is the scheduling action. Execute scheduling based on scheduling actions; The enhanced status includes predicted data and the current real-time status of the system; the predicted data includes the predicted downlink data volume and task execution duration, the corresponding CPU core requirements and memory resource requirements; Before the execution scheduling, a resource adaptation check is performed. If resources are sufficient, they are allocated directly. If resources are insufficient and it is a high-priority task, a preemption mechanism is triggered to preempt resources from low-priority regular tasks. If there are no preemption conditions, the action is invalid.

2. The computational resource scheduling method for ground data processing operations of space science satellites according to claim 1, characterized in that, The initial values ​​of the parameters of the deep reinforcement learning model are obtained by pre-training the reinforcement learning policy network using an expert scheduling strategy, including: Shortest job priority is selected as the expert strategy. Full scheduling is performed on the simulation task set, and the state-action trajectory of each step is recorded to form an expert trajectory set. A convolutional neural network is used as the imitation learning model, with the system state as input and expert actions as supervision labels. The network is trained using mini-batch stochastic gradient descent and cross-entropy loss function.

3. The computational resource scheduling method for ground data processing operations of space science satellites according to claim 1, characterized in that, The process of obtaining the prediction data includes: Input the historical task time-series feature vectors into the trained LSTM network to output predicted data; The historical task time-series feature vector includes time embedding features, historical downlink data volume, execution duration, payload type, and task priority.

4. The computational resource scheduling method for ground data processing operations of space science satellites according to claim 1, characterized in that, The current real-time status of the system is as follows: The system state is encoded into a multi-channel two-dimensional matrix, which is divided into three parts: The cluster resource allocation area is used to represent the resource usage and task characteristics at each time step; The waiting queue area is used to represent the resource requirements, duration, and priority of tasks to be scheduled; The backlog queue area represents the number of tasks waiting to be scheduled that exceeds the capacity of the waiting queue.

5. The computational resource scheduling method for ground data processing operations of space science satellites according to claim 1, characterized in that, The reward function of the deep reinforcement learning model is: ； Where λ is the resource utilization rate weight; κ is the priority intensity coefficient; For the task Priority; For the task The time already spent in the area, For the task arrival time, For a moment System time; U(t) This represents the average resource utilization rate of the cluster. For a moment The set of all incomplete tasks in the system.

6. The computational resource scheduling method for ground data processing operations of space science satellites according to claim 1, characterized in that, During the operation of the deep reinforcement learning model, the policy network and value network are iteratively updated, including: Collect trajectory data; Calculate the cumulative discount reward and generalized advantage estimate, and update the value network; The policy network is updated based on the pruning objective function, and the interaction, sampling and updating are performed cyclically until the policy converges.

7. A computational resource scheduling system for ground data processing operations of space science satellites, implemented based on the method described in any one of claims 1-6, characterized in that, The system includes: The scheduling module is used to take the augmentation state as input and output the scheduling action, with the proximal policy optimization algorithm as the core algorithm of the deep reinforcement learning model. The execution scheduling module is used to perform scheduling based on scheduling actions; before executing the scheduling, a resource adaptation check is performed, and if resources are sufficient, they are allocated directly; if resources are insufficient and it is a high-priority task, a preemption mechanism is triggered to preempt resources from low-priority regular tasks; if there are no preemption conditions, the action is invalid; and An augmented state module is constructed to build the augmented states required for the input of the deep reinforcement learning model. The augmented states include the predicted data and the current real-time state of the system. The predicted data includes the predicted downlink data volume and task execution duration, the corresponding CPU core requirements, and memory resource requirements.

Images