Power logistics operation decoupling method, system and device based on reinforcement learning
By constructing a global coupling graph and performing graph decomposition based on reinforcement learning, and using multi-agent reinforcement learning to optimize cross-field group tasks, the problem of heterogeneous processing of task data and coupling relationships in power logistics systems is solved, thereby improving the adaptability and execution efficiency of the scheduling system.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-25
- Publication Date
- 2026-03-10
AI Technical Summary
Existing power logistics systems suffer from several problems in cross-field group operations, including insufficient ability to uniformly process heterogeneous task data, difficulty in accurately characterizing and generalizing task coupling relationships, weak ability of scheduling models to optimize multi-task collaboration, and lack of dynamic feedback and closed-loop optimization mechanisms.
We adopt a reinforcement learning-based approach, which involves preprocessing task data, constructing a global coupling graph and performing graph decomposition, and using multi-agent reinforcement learning to achieve cross-subgraph collaborative optimization. We design a multi-agent reinforcement learning architecture with collaborative awareness and introduce a closed-loop optimization mechanism.
It achieves unified modeling and feature extraction of heterogeneous task data, improves scheduling optimization capabilities and task coverage, enhances the system's adaptability and execution efficiency in complex scenarios, and improves the adaptive adjustment capability of scheduling strategies.
Smart Images
Figure CN121639062A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the field of intelligent scheduling and power logistics optimization, and in particular to a power logistics operation decoupling method, system and device based on reinforcement learning. BACKGROUND
[0002] With the continuous development of power logistics systems and intelligent scheduling technology, a large amount of heterogeneous information is generated in daily power operation and material transportation for cross-site group operation tasks, covering different types of task data such as equipment maintenance tasks, material transportation tasks, emergency repair tasks, cross-site vehicle scheduling tasks, and resource state data such as transportation lines, storage capacity, vehicles and lifting equipment. At present, the mainstream cross-site group operation scheduling method is still dominated by manual experience or static rules, combined with limited optimization models for preliminary scheduling.
[0003] Specific problems mainly include:
[0004] 1) Insufficient ability to process heterogeneous task data. Cross-site group operation tasks involve time series information, geographic location, resource state and operation priority, and there are significant differences between different sources. The existing method lacks unified modeling and fusion mechanism, which makes it difficult to fully express the operation state characteristics and limits the depth and accuracy of scheduling optimization.
[0005] 2) It is difficult to accurately describe and generalize the task coupling relationship. Traditional methods rely on manual experience or static rules to judge task dependencies, and lack general coupling relationship modeling and graph structure expression ability. In the case of diverse task types or frequent changes in operation scenarios, the system is difficult to quickly adapt, resulting in scheduling coverage blind area or resource conflict.
[0006] 3) Weak multi-task coordination optimization capability of scheduling model. Cross-site group operation often includes transportation, installation, maintenance and other tasks, and there are significant differences in time, space and resource coupling between different tasks. It is difficult for a general optimization model to consider all task objectives, resulting in low scheduling efficiency or frequent task execution conflicts.
[0007] 4) Lack of effective dynamic feedback and closed-loop optimization mechanism. Most existing scheduling systems rely on static rules or pre-planned task sequences, lack intelligent strategy adjustment capability based on real-time operation state, and cannot realize multi-agent collaborative learning and adaptive optimization. The efficiency and stability of the scheduling system decrease with the increase of task size and complexity. SUMMARY
[0008] In view of the problems existing in the prior art, the present application discloses a power logistics operation decoupling method, system and device based on reinforcement learning.
[0009] A power logistics operation decoupling method based on reinforcement learning, comprising the following steps:
[0010] Obtaining task data in power logistics operation, preprocessing the task data to obtain standardized task vectors and joint modeling;
[0011] Taking the standardized task vectors as initial representations of nodes, constructing a global coupling relationship graph of power logistics cross-site tasks;
[0012] Converting the global coupling relationship graph into a sparse graph, performing a graph decomposition operation on the sparse graph to obtain N subgraphs;
[0013] Implementing cross-subgraph collaborative optimization using a multi-agent reinforcement learning method to obtain collaborative scheduling of cross-site group tasks;
[0014] Feeding back the collaborative scheduling of cross-site group tasks to the power logistics cross-site group intelligent scheduling system.
[0015] As an implementable manner, the standardized task vectors include task description vectors, resource state vectors, and timing embedding vectors.
[0016] As an implementable manner, obtaining task data in power logistics operation, preprocessing the task data to obtain standardized task vectors includes the following steps:
[0017] Collecting task data in power logistics operation and converting the collected data into task description vectors;
[0018] Obtaining state information of vehicles, equipment, and manpower from a resource management system to form resource state vectors;
[0019] Building historical task execution records and resource utilization into timing sequences, extracting dynamic features through a time series analysis function to obtain timing embedding vectors;
[0020] In the fusion encoding layer, joint modeling of the task description vectors, resource state vectors, and timing embedding vectors.
[0021] As an implementable manner, taking the standardized task vectors as initial representations of nodes, constructing a global coupling relationship graph of power logistics cross-site tasks includes the following steps:
[0022] Based on the standardized task vectors as initial representations of nodes, constructing a global coupling relationship graph of power logistics cross-site tasks;
[0023] After the coupling relationship graph is constructed, iterative updating of node representations is performed using a graph neural network to capture high-order dependency relationships of cross-site power logistics tasks and form a global representation of the overall task network.
[0024] As one possible implementation, the step of converting the global coupling graph into a sparse graph and performing a graph decomposition operation on the sparse graph to obtain N subgraphs includes the following steps:
[0025] For any two task nodes, calculate the overall similarity between the nodes;
[0026] Task nodes with a comprehensive similarity value lower than a preset threshold are removed from the graph, resulting in a sparse graph with weakly correlated edges removed.
[0027] The system performs a graph decomposition operation on the sparse graph to obtain N subgraphs.
[0028] As one possible implementation method, the cross-subgraph cooperative scheduling using a multi-agent reinforcement learning method includes the following steps:
[0029] Define the state space and action space;
[0030] The multi-agent architecture adopts a centralized training and distributed execution architecture. During the training phase, it shares global state information and optimizes the cooperative strategy through policy gradient or value distribution methods, assigning tasks within each subgraph to a group of agents.
[0031] As one possible implementation method, the advantage function used in the multi-agent reinforcement learning method is expressed as:
[0032]
[0033] Where At represents the current task cluster scheduling state advantage, Rt represents the delay penalty at the current time t, and V(s) t V(s) represents the current state value. t+1 ) represents the value of the next state, and γ represents the discount factor.
[0034] A reinforcement learning-based decoupling system for power logistics operations includes:
[0035] The task data acquisition module is used to acquire task data in power logistics operations, preprocess the task data to obtain standardized task vectors, and perform joint modeling.
[0036] The coupling graph construction module is used to construct a global coupling graph of cross-site tasks in power logistics by using the standardized task vector as the initial representation of the nodes.
[0037] The task decoupling and decomposition module is used to convert the global coupling relationship graph into a sparse graph, and perform graph decomposition operation on the sparse graph to obtain N subgraphs.
[0038] The multi-agent scheduling module is used to achieve cross-subgraph collaborative optimization by utilizing multi-agent reinforcement learning methods, thereby obtaining the collaborative scheduling of cross-field group tasks;
[0039] The feedback update module is used to update the collaborative scheduling feedback of cross-field group tasks to the intelligent scheduling system for cross-field group power logistics.
[0040] A computer-readable storage medium storing a computer program that, when executed by a processor, implements the following method:
[0041] Acquire task data from power logistics operations, preprocess the task data to obtain standardized task vectors, and perform joint modeling.
[0042] Using the standardized task vector as the initial representation of the node, a global coupling relationship graph of cross-site tasks in power logistics is constructed;
[0043] The global coupling graph is converted into a sparse graph, and a graph decomposition operation is performed on the sparse graph to obtain N subgraphs;
[0044] By utilizing multi-agent reinforcement learning methods to achieve cross-subgraph collaborative optimization, collaborative scheduling of cross-field group tasks is obtained;
[0045] The collaborative scheduling feedback of cross-site group tasks will be updated to the intelligent scheduling system for cross-site groups of power logistics.
[0046] A power logistics operation decoupling device based on reinforcement learning includes a memory, a processor, and a computer program stored in the memory and running on the processor. When the processor executes the computer program, it implements the following method:
[0047] Acquire task data from power logistics operations, preprocess the task data to obtain standardized task vectors, and perform joint modeling.
[0048] Using the standardized task vector as the initial representation of the node, a global coupling relationship graph of cross-site tasks in power logistics is constructed;
[0049] The global coupling graph is converted into a sparse graph, and a graph decomposition operation is performed on the sparse graph to obtain N subgraphs;
[0050] By utilizing multi-agent reinforcement learning methods to achieve cross-subgraph collaborative optimization, collaborative scheduling of cross-field group tasks is obtained;
[0051] The collaborative scheduling feedback of cross-site group tasks will be updated to the intelligent scheduling system for cross-site groups of power logistics.
[0052] This invention, by adopting the above technical solutions, has significant technical effects:
[0053] 1) This invention supports unified modeling and feature extraction of heterogeneous task data, realizes standardized representation of job status and deep integration of resource information, and improves scheduling optimization capability and task coverage;
[0054] 2) This invention constructs a highly scalable and generalizable task dependency expression method through task coupling relationship graph and graph decomposition mechanism, which significantly improves the system's adaptability to complex operation scenarios;
[0055] 3) This invention designs a multi-agent reinforcement learning architecture with collaborative perception capabilities, which realizes dynamic assignment, resource coordination and task collaborative optimization of cross-field group tasks, and enhances the system's execution efficiency and stability under multi-objective scheduling.
[0056] 4) This invention introduces a closed-loop optimization mechanism based on task execution status and feedback to achieve adaptive adjustment of scheduling strategies and improve the robustness and global optimization capability of the system in dynamic scenarios. Attached Figure Description
[0057] Figure 1 This is a flowchart illustrating the reinforcement learning-based power logistics operation decoupling method described in this invention.
[0058] Figure 2 This is an overall schematic diagram of the power logistics operation decoupling system based on reinforcement learning described in this invention. Detailed Implementation
[0059] The present invention will be further described below with reference to the accompanying drawings and embodiments. The following embodiments are explanations of the present invention, but the present invention is not limited to the following embodiments.
[0060] Example 1:
[0061] A digital twin-based method for optimizing the scheduling of power logistics depot clusters, such as... Figure 1 As shown, it includes the following steps:
[0062] S100. Obtain task data in power logistics operations, preprocess the task data to obtain standardized task vectors and perform joint modeling.
[0063] S200. Using the standardized task vector as the initial representation of the node, construct a global coupling relationship graph for cross-site tasks in power logistics;
[0064] S300. Convert the global coupling graph into a sparse graph, and perform a graph decomposition operation on the sparse graph to obtain N subgraphs.
[0065] S400: Utilize multi-agent reinforcement learning methods to achieve cross-subgraph collaborative optimization, thereby obtaining collaborative scheduling of cross-field group tasks;
[0066] S500 updates the collaborative scheduling feedback of cross-field group tasks to the power logistics cross-field group intelligent scheduling system.
[0067] In this embodiment, step S100 involves acquiring task data from power logistics operations, preprocessing the task data to obtain standardized task vectors, and performing joint modeling. The standardized task vectors include task description vectors, resource state vectors, and temporal embedding vectors. The steps of acquiring task data from power logistics operations and preprocessing the task data to obtain standardized task vectors include the following steps:
[0068] S110: Collect task data in power logistics operations and convert the collected data into task description vectors.
[0069] Collect raw data on various types of tasks, including equipment maintenance, material transportation, emergency repair, and cross-site vehicle dispatching. Combine this data with the status of vehicles, equipment, and human resources to establish status vectors for availability and capacity.
[0070] S120: Obtain the status information of vehicles, equipment and personnel from the resource management system to form a resource status vector.
[0071] S130: Construct a time series of historical task execution records and resource utilization information, extract dynamic features through time series analysis functions, and obtain a time series embedding vector.
[0072] S140: In the fusion coding layer, the task description vector, resource state vector, and temporal embedding vector are jointly modeled to obtain a unified feature vector. The unified feature vector is obtained as follows:
[0073]
[0074] Among them, T i Let x represent the unified feature vector of task i. i r represents the task description vector. i h represents the resource state vector. i This represents a temporal embedding vector.
[0075] In this embodiment, step S200, which involves using the standardized task vector as the initial representation of the node to construct a global coupling relationship graph for cross-site tasks in power logistics, includes the following steps:
[0076] S210: Based on the standardized task vector as the initial representation of the node, construct a global coupling relationship graph for cross-site tasks in power logistics;
[0077] S220: After the coupling relationship graph is constructed, the graph neural network is used to iteratively update the node representation to capture the high-order dependencies of cross-field power logistics tasks and form a global representation of the overall task network.
[0078] In this embodiment, step S300, which involves converting the global coupling graph into a sparse graph and performing a graph decomposition operation on the sparse graph to obtain N subgraphs, includes the following steps:
[0079] S310: For any two task nodes, calculate the overall similarity between the nodes;
[0080] S320: Remove task nodes with a comprehensive similarity value lower than a preset threshold from the graph to obtain a sparse graph with weakly correlated edges removed.
[0081] S330: The system performs a graph decomposition operation on the sparse graph to obtain N subgraphs.
[0082] In this embodiment, step S400, which utilizes a multi-agent reinforcement learning method to achieve cross-subgraph cooperative scheduling, includes the following steps:
[0083] S410: Define the state space and action space. After task decomposition, design a multi-agent reinforcement learning framework to achieve collaborative scheduling within and across task clusters. Assign tasks within each subgraph to a group of agents, with each agent corresponding to a type of execution resource.
[0084] Define a state space S that includes task progress, resource status, job location, etc.; and an action space A that includes operations such as task allocation, path selection, and time scheduling.
[0085] S420: The multi-agent architecture adopts a centralized training and distributed execution architecture. During the training phase, it shares global state information and optimizes the cooperative strategy through policy gradient or value distribution methods, assigning tasks within each subgraph to a group of agents.
[0086] A multi-objective function that comprehensively considers task completion rate, transportation costs, energy consumption, and delay penalties:
[0087]
[0088] Where α, β, γ, δ represent adjustable parameters, R represents the reward function, Comp represents the task completion rate, Cost represents the transportation cost, Delay represents the delay penalty, and Energy represents the source consumption.
[0089] The multi-agent architecture employs centralized training and distributed execution. During training, it shares global state information and optimizes cooperative strategies using policy gradients or value distribution methods. During execution, each agent makes autonomous decisions based on local information, achieving efficient collaborative scheduling across tasks. The reinforcement learning network as a whole adopts an internal agent-value function network architecture. The internal agents take the executable resource state as input and output an action probability distribution, while the value function network takes the executable resource state as input and outputs a state value. During the internal agent's execution of network parameter updates, it first calculates the advantage function.
[0090] The advantage function used in the multi-agent reinforcement learning method is expressed as follows:
[0091]
[0092] Where At represents the current task cluster scheduling state advantage, Rt represents the delay penalty at the current time t, and V(s) t V(s) represents the current state value. t+1 ) represents the value of the next state, and γ represents the discount factor.
[0093] After obtaining the advantage At, the Actor network optimizes the policy parameters through gradient ascent, and the optimization process can be expressed as:
[0094]
[0095] Where Δθ represents the Actor network parameter update amount, η represents the learning rate, θ represents the Actor network parameters, and s t This indicates the current state, including task information and resource status within and across task clusters. t The current action, including task allocation, path selection, and time scheduling, is represented by a policy probability distribution, A. t This represents the advantage function, and the method is used to make the action selection probability related to A. t The direction is consistent, thereby improving the scheduling effect within and across task clusters.
[0096] In this embodiment, step S500 involves updating the collaborative scheduling feedback of cross-field group tasks to the power logistics cross-field group intelligent scheduling system.
[0097] This invention supports unified modeling and feature extraction of heterogeneous task data, achieving standardized representation of job status and deep integration of resource information, thereby improving scheduling optimization capabilities and task coverage. Through task coupling graphs and graph decomposition mechanisms, it constructs a highly scalable and generalizable task dependency representation, significantly enhancing the system's adaptability to complex job scenarios. A multi-agent reinforcement learning architecture with collaborative awareness is designed to achieve dynamic assignment, resource coordination, and task collaborative optimization across different task groups, enhancing the system's execution efficiency and stability under multi-objective scheduling. A closed-loop optimization mechanism based on task execution status and feedback is introduced to achieve adaptive adjustment of scheduling strategies, improving the system's robustness and global optimization capabilities in dynamic scenarios.
[0098] Example 2:
[0099] A power logistics operation decoupling system based on reinforcement learning includes: a task data acquisition module 100, a coupling relationship graph construction module 200, a task decoupling and decomposition module 300, a multi-agent scheduling module 400, and a feedback update module 500.
[0100] The task data acquisition module 100 is used to acquire task data in power logistics operations, preprocess the task data to obtain standardized task vectors, and perform joint modeling.
[0101] The coupling graph construction module 200 is used to construct a global coupling graph of cross-field tasks in power logistics by using the standardized task vector as the initial representation of the node.
[0102] The task decoupling and decomposition module 300 is used to convert the global coupling relationship graph into a sparse graph, and perform graph decomposition operation on the sparse graph to obtain N subgraphs.
[0103] The multi-agent scheduling module 400 is used to achieve cross-subgraph collaborative optimization using multi-agent reinforcement learning methods, thereby obtaining the collaborative scheduling of cross-field group tasks.
[0104] The feedback update module 500 is used to update the collaborative scheduling feedback of cross-field group tasks to the intelligent scheduling system for cross-field group power logistics.
[0105] Various changes and modifications made without departing from the spirit and scope of this invention, and all equivalent technical solutions, also fall within the scope of this invention.
[0106] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.
[0107] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, apparatus, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.
[0108] This invention is described with reference to flowchart illustrations and / or block diagrams of the method, terminal device (system), and computer program product according to the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing terminal device to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing terminal device, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.
[0109] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing terminal device to operate in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.
[0110] These computer program instructions can also be loaded onto a computer or other programmable data processing terminal equipment, causing a series of operational steps to be performed on the computer or other programmable terminal equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable terminal equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.
[0111] It should be noted that:
[0112] The phrase "an embodiment" or "an embodiment" used in this specification means that a particular feature, structure, or characteristic described in connection with the embodiment is included in at least one embodiment of the invention. Therefore, the phrase "an embodiment" or "an embodiment" appearing in various places throughout the specification does not necessarily refer to the same embodiment.
[0113] Furthermore, it should be noted that the shapes and names of the components in the specific embodiments described in this specification may differ. All equivalent or simple variations made to the structure, features, and principles described in this patent concept are included within the protection scope of this patent. Those skilled in the art to which this invention pertains may make various modifications or additions to the described specific embodiments or use similar methods to substitute them, as long as they do not depart from the structure of this invention or exceed the scope defined in these claims, they should all fall within the protection scope of this invention.
Claims
1. A power logistics operation decoupling method based on reinforcement learning, characterized in that, The method comprises the following steps: acquiring task data in power logistics operation, preprocessing the task data to obtain standardized task vectors and performing joint modeling; using the standardized task vectors as initial representations of nodes to construct a global coupling relationship graph of power logistics cross-site tasks; converting the global coupling relationship graph into a sparse graph, performing graph decomposition operation on the sparse graph to obtain N subgraphs; using a multi-agent reinforcement learning method to realize cross-subgraph collaborative optimization to obtain collaborative scheduling of cross-site group tasks; feeding back the collaborative scheduling of cross-site group tasks to the power logistics cross-site group intelligent scheduling system.
2. The power logistics operation decoupling method based on reinforcement learning according to claim 1, characterized in that, The standardized task vectors include task description vectors, resource state vectors and timing embedding vectors. 3.The power logistics operation decoupling method based on reinforcement learning according to claim 2, characterized in that, The acquiring task data in power logistics operation, preprocessing the task data to obtain standardized task vectors comprises the following steps: collecting task data in power logistics operation and converting the collected data into task description vectors; obtaining state information of vehicles, equipment and manpower from a resource management system to form resource state vectors; constructing historical task execution records and resource utilization into timing sequences, extracting dynamic characteristics through a time sequence analysis function to obtain timing embedding vectors; jointly modeling the task description vectors, resource state vectors and timing embedding vectors in a fusion coding layer.
4. The power logistics operation decoupling method based on reinforcement learning according to claim 1, characterized in that, The using the standardized task vectors as initial representations of nodes to construct a global coupling relationship graph of power logistics cross-site tasks comprises the following steps: based on the standardized task vectors as initial representations of nodes, constructing a global coupling relationship graph of power logistics cross-site tasks; after the coupling relationship graph is constructed, iteratively updating the node representations by using a graph neural network to capture high-order dependency relationships of cross-site power logistics tasks and form a global representation of the overall task network.
5. The method of claim 1, wherein, The converting the global coupling relationship graph into a sparse graph, performing graph decomposition operation on the sparse graph to obtain N subgraphs comprises the following steps: calculating the comprehensive similarity between nodes for any two task nodes; removing task nodes with a comprehensive similarity value lower than a preset threshold from the graph to obtain a sparse graph with weakly related edges removed; the system performs graph decomposition operation on the sparse graph to obtain N subgraphs.
6. The power logistics operation decoupling method based on reinforcement learning according to claim 1, characterized in that, The using a multi-agent reinforcement learning method to realize cross-subgraph collaborative optimization comprises the following steps: defining a state space and an action space; the multi-agent adopts a centralized training and distributed execution architecture, shares global state information in the training stage, optimizes the collaboration strategy by using a policy gradient or value distribution method, and allocates tasks within each subgraph to a group of agents.
7. The power logistics operation decoupling method based on reinforcement learning according to claim 6, wherein the advantage function used in the multi-agent reinforcement learning method is represented as: Where At represents the current task cluster scheduling state advantage, Rt represents the delay penalty at the current time t, and V(s) t V(s) represents the value of the current state. t+1 ) represents the value of the next state, and γ represents the discount factor.
8. A power logistics operation decoupling system based on reinforcement learning, characterized in that, including: a task data collection module for acquiring task data in power logistics operation, preprocessing the task data to obtain standardized task vectors and performing joint modeling; a coupling relationship graph construction module for using the standardized task vectors as initial representations of nodes to construct a global coupling relationship graph of power logistics cross-site tasks; and a collaborative scheduling module for using a multi-agent reinforcement learning method to realize cross-subgraph collaborative optimization to obtain collaborative scheduling of cross-site group tasks. A task decoupling and decomposition module is configured to convert a global coupling graph into a sparse graph, perform a graph decomposition operation on the sparse graph, and obtain N subgraphs; A multi-agent scheduling module is configured to implement cross-subgraph collaborative optimization by using a multi-agent reinforcement learning method, and obtain collaborative scheduling of cross-field group tasks. A feedback updating module is configured to feed back the collaborative scheduling of the cross-field group tasks to the power logistics cross-field group intelligent scheduling system.
9. A computer-readable storage medium storing a computer program, the computer program comprising instructions that, when executed by a computer, cause the computer to perform the method of any one of claims 1 to 8. The computer program, when executed by a processor, implements the method according to any one of claims 1 to 7.
10. A power logistics operation decoupling device based on reinforcement learning, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, The processor, when executing the computer program, implements the method according to any one of claims 1 to 7.