A communication base station energy storage array dynamic reconstruction method and system based on multi-dimensional working condition coupling

By using a distributed multi-agent architecture and the Transformer-XL model, the problems of high scheduling complexity and unreasonable resource allocation in the energy storage system of communication network base stations are solved, realizing collaborative optimization and efficient resource management among base stations, and improving the robustness and survivability of the system.

CN123070909APending Publication Date: 2026-11-10HAIJIU LIULING INTERNET CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611090316.3
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-07-22
Publication Date
2026-11-10

AI Technical Summary

Technical Problem

Existing communication network base station energy storage systems suffer from high computational complexity in centralized scheduling, inability to handle long-term hidden battery damage, and lack of refined supply and demand matching mechanisms, resulting in low overall network survivability.

Method used

A dynamic reconfiguration method for communication base station energy storage arrays based on multi-dimensional operating condition coupling is adopted. By utilizing a distributed multi-agent architecture and the Transformer-XL time series model, a virtual bidding strategy is generated through self-awareness embedding and environmental awareness embedding, thereby realizing collaborative decision-making and resource optimization among base stations.

Benefits of technology

It significantly reduces computational complexity, improves system robustness, achieves efficient optimization of network resources and enhances global survivability, and avoids resource waste and single point of failure risks.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123070909A_ABST
    Figure CN123070909A_ABST
Patent Text Reader

Abstract

The application provides a communication base station energy storage array dynamic reconstruction method and system based on multi-dimensional working condition coupling, belongs to the field of communication network operation and maintenance, and is used for solving the problems of high complexity of centralized scheduling, inability to evaluate long-time hidden damage of batteries and non-optimal resource configuration. Each base station and its energy storage unit are modeled as a reinforcement learning intelligent agent with a virtual account; each intelligent agent processes battery historical data by using a Transformer-XL model to generate a self-cognition embedding vector; the intelligent agent broadcasts and aggregates encrypted intention feature vectors through a graph network to form an environment perception embedding vector; the two vectors are combined to generate a buy or sell offer for asset replacement; a centralized training and distributed execution architecture is used to simulate multiple rounds of bidding, and physical battery reconstruction, digital twin data migration and model memory hot loading are synchronously performed. The application adopts a distributed intelligent agent architecture, the calculation complexity linearly grows with the scale, and has high expansibility and robustness.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention belongs to the field of communication network infrastructure operation and maintenance technology, specifically involving a dynamic reconfiguration method and system for communication base station energy storage arrays based on multi-dimensional operating condition coupling. Background Technology

[0002] With the explosive growth in 5G / 6G base station density and the surge in power consumption per station, the energy security of communication networks faces severe challenges. Base station backup batteries are not only the last line of defense against mains power outages but also highly valuable dispatchable assets. However, in the existing base station energy storage operation and maintenance system, the allocation and scheduling of resources suffer from the following fundamental technical bottlenecks: 1. The "Curse of Dimensionality" and Single Point of Failure Risk of Centralized Scheduling: Existing technologies mostly employ centralized control centers for unified scheduling across the entire network. Facing tens of thousands of base station nodes, optimizing the allocation of energy storage assets presents a significant challenge. The existing passive maintenance model of "replacing the faulty one" lacks a global asset flow mechanism, resulting in high-value (high SOH) batteries remaining idle at low-risk sites for extended periods, while high-risk sites, lacking high-quality batteries, are the first to fail in power outages, leading to a severe waste of network-wide asset utility. Furthermore, if the central node experiences communication failure or computing power paralysis, network-wide scheduling will instantly fail, resulting in extremely poor system robustness.

[0003] 2. The "Black Box" of Battery Aging Status and Short-Sighted Matching: Traditional scheduling simply matches batteries based on their current state of health (SOH) or voltage value, creating a weak-strong pairing. However, electrochemical batteries exhibit significant non-Markovian aging characteristics, meaning their current performance is deeply influenced by historical operating conditions from several years ago (such as a deep over-discharge or high-temperature exposure). Existing scheduling algorithms lack the ability to extract this long-term memory, leading to the placement of potentially faulty batteries ("strong on the outside but weak on the inside") at high-risk base stations, creating a risk of battery dropout.

[0004] 3. Lack of incentive-compatible coordination mechanisms: In actual operation and maintenance, base stations in different areas often belong to different maintenance grids or even different operators (in shared tower scenarios). Traditional mandatory command-based scheduling ignores local interests and lacks a mechanism to measure "which base station needs this battery most at this moment." The lack of a dynamic pricing mechanism leads to resources often flowing to nodes that are not in the most urgent need, resulting in a waste of the overall network's survivability. Summary of the Invention

[0005] In view of this, the present invention provides a method and system for dynamic reconfiguration of energy storage arrays for communication base stations based on multi-dimensional operating condition coupling, in order to solve the technical problems in the prior art, such as high computational complexity of centralized scheduling, inability to handle long-term hidden damage of batteries, and low network survivability caused by lack of a refined supply and demand matching mechanism.

[0006] The technical solution adopted in this invention is as follows: A dynamic reconfiguration method for communication base station energy storage arrays based on multi-dimensional operating condition coupling includes: Step 1: Model each communication base station and its associated energy storage unit in the target area as an independent reinforcement learning agent, and associate each agent with the corresponding base station operation requirements and energy storage unit status; In step 1, each of the intelligent agents is configured with a local observation space and a virtual utility account; The local observation space is used to observe the local load of the corresponding base station, the real-time voltage of the auxiliary energy storage unit, the operating environment temperature, and the historical operation and maintenance logs of both. The virtual utility account is used to pay or receive virtual credit points during the bidding process, quantifying the urgency of the corresponding base station's demand for high-quality energy storage resources.

[0007] Step 2: Each agent processes the historical lifecycle operation data of the energy storage unit based on the Transformer-XL time series model to generate a self-awareness embedding vector that includes the battery's hidden damage memory and future performance degradation trend; Step 2 specifically includes: Step 2.1: Each intelligent agent acquires time-series operating data of its energy storage unit during its historical life cycle, the data including at least voltage, current and temperature sequences; Step 2.2: Input the historical time series runtime data into the locally deployed Transformer-XL time series model; Step 2.3: The Transformer-XL time series model processes the historical time series running data and finally outputs the self-awareness embedding vector; In the process of calculating the self-attention score in the Transformer-XL model, a temperature control factor weighted attention mechanism is applied to calculate a bias weight based on temperature records for each historical time segment and add this bias weight to the calculation of the attention score.

[0008] The calculation formula for the temperature control factor weighted attention mechanism is as follows:

[0009] Where Q is the query vector, K is the key vector, and V is the value vector. The dimension of the key vector. This is the transpose operation of the key vector K, and softmax is the normalization function. For temperature control bias matrix, The calculation method is as follows:

[0010] in, Representing historical moments Battery temperature records, Indicates the battery's safe temperature threshold. This represents the thermal damage memory coefficient, where t is the current time step and j is the historical time step. This is the temperature control bias value of the current time step t relative to the historical time step j.

[0011] Step 3: Each agent constructs a dynamic topology graph through the communication links between its base stations, broadcasts encrypted intent feature vectors, and aggregates neighbor features using the intent-aware graph network to form an environmental perception embedding vector that characterizes the resonance level of regional power supply risk. Step 3 specifically includes the following steps: Step 3.1: Each agent constructs a dynamic topology graph between base stations based on the geographical proximity and power supply topology of its base station, and dynamically maintains a list of one-hop neighbor nodes; Step 3.2: Each agent inputs its own base station operation requirements and energy storage unit status into a lightweight MLP encoder to generate an encrypted semantic intent feature vector containing both the request for help intent and the cooperation intent; Step 3.3: Broadcast the intent feature vector to neighboring nodes through the inter-base station communication link, without transmitting the original operation data; Step 3.4: Calculate the attention coefficients of the intent feature vectors of neighboring nodes using the graph attention mechanism, aggregate the neighboring features according to the coefficient weights, and generate the environment-aware embedding vector.

[0012] In step 3.4, the environment-aware embedding vector Calculated using the following formula:

[0013] in, These are the attention weight coefficients of the intention feature vector of neighboring agent j. W It is a shared weight matrix. It is the intention feature vector of neighboring agent j. It is the activation function, N( i ) is the intelligent agent i i The set of neighbors.

[0014] Step 4: Each agent combines its self-awareness embedding vector and environmental perception embedding vector, and integrates its own base station operation requirements and energy storage unit status to autonomously generate a buy or sell price for asset swap, forming a virtual bidding strategy based on risk pricing. In step 4, the generation of the virtual bidding strategy follows the principle of diminishing marginal utility, specifically: When an agent predicts that the current battery health degradation will lead to a future power supply capacity gap, it generates a buy offer based on the marginal utility of the expected increase in lifespan; the growth rate of the buy offer slows down as the offer increases, and the upper limit of the offer is limited by the credit score balance in the agent's virtual account; When the agent determines that the current battery health is sufficient and the risk to itself after the sale is controllable, it generates a selling price based on the reduction in its own survival utility caused by the sale of the battery; the selling price is the lowest acceptable price that the agent is willing to accept for the sale of its battery entity.

[0015] Step 5: Simulate multiple rounds of asset swap bidding by training a policy network with a distributed execution architecture in a centralized manner. Search for a Nash equilibrium strategy combination that maximizes the long-term survival probability of the entire network and minimizes logistics entropy. After reaching an asset swap contract that satisfies cost constraints, simultaneously execute the physical reconstruction of energy storage units, the on-path migration of digital twins, and the hot loading of model memory.

[0016] Step 5, the model memory hot loading specifically includes: The target base station agent receives and decrypts data packets from the source base station; The decrypted data is loaded into its built-in Transformer-XL time series model to update the model's historical memory of the current physical battery; based on the updated memory, the battery's health status and degradation trend are continuously evaluated.

[0017] In step 5, the implementation of the centralized training distributed execution architecture specifically includes: During the offline training phase, a central critic network is constructed, which receives state information from all agents and calculates the global value function and advantage function. Using the global advantage function, the local policy networks of all agents are trained simultaneously. During the online execution phase, each agent uses only the trained local policy network to make independent decisions based on its own and its neighbors' information.

[0018] Example 2

[0019] This embodiment proposes a distributed collaborative scheduling system for energy storage resources of communication base stations, the system comprising: Distributed intelligent agent nodes are deployed at various communication base stations to perform local state perception, cognitive reasoning, and bidding decision-making. The graph communication consensus module is used to maintain dynamic neighbor relationships between base stations and to handle the broadcasting and aggregation of intent feature vectors. The virtual auction matching engine, deployed in the cloud, is used to receive and match buy and sell quotes, verify cost constraints, and generate scheduling instructions. The policy evolution training platform, deployed on a server, is used to perform centralized reinforcement learning training and update and distribute policy parameters to each agent node.

[0020] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are: 1. This invention employs a distributed multi-agent architecture, with decision-making computations distributed across the edges of each base station. As the number of base stations increases, the computational complexity grows linearly rather than exponentially, exhibiting strong scalability. Furthermore, the failure of any node will not lead to network-wide paralysis, significantly improving the system's robustness.

[0021] 2. This invention introduces virtual bidding and Nash equilibrium game, which not only solves the priority problem of resource allocation (i.e. "who needs it most"), but also automatically balances the high costs of physical transportation and manual on-site installation with the survival benefits of base stations through economic mechanisms. This avoids ineffective asset rotation caused by short-sighted local perspectives and ensures that every physical reconstruction brings a significant improvement in global reliability. Attached Figure Description

[0022] The present invention will be described by way of example and with reference to the accompanying drawings, wherein: Figure 1 This is a diagram of the overall architecture of the present invention. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0024] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0025] It should be noted that, unless otherwise specified, the embodiments and features described in this invention can be combined with each other.

[0026] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0027] In this invention, unless otherwise explicitly specified and limited, "above" or "below" the second feature can include direct contact between the first and second features, or contact between the first and second features through another feature between them. Furthermore, "above," "over," and "on top" of the second feature includes the first feature directly above or diagonally above the second feature, or simply indicates that the first feature is at a higher horizontal level than the second feature. "Below," "below," and "under" the second feature includes the first feature directly below or diagonally below the second feature, or simply indicates that the first feature is at a lower horizontal level than the second feature.

[0028] It should be noted that, unless otherwise specified, the embodiments and features described in this invention can be combined with each other.

[0029] Example 1

[0030] like Figure 1 As shown in the figure, an embodiment of the present invention discloses a dynamic reconfiguration method for a communication base station energy storage array based on multi-dimensional operating condition coupling, comprising: Step 1: Model each communication base station and its associated energy storage unit in the target area as an independent reinforcement learning agent, and associate each agent with the corresponding base station operation requirements and energy storage unit status; In step 1, each of the intelligent agents is configured with a local observation space and a virtual utility account; The local observation space is used to observe the local load of the corresponding base station, the real-time voltage of the auxiliary energy storage unit, the operating environment temperature, and the historical operation and maintenance logs of both. The virtual utility account is used to pay or receive virtual credit points during the bidding process, quantifying the urgency of the corresponding base station's demand for high-quality energy storage resources.

[0031] Step 2: Each agent processes the historical lifecycle operation data of the energy storage unit based on the Transformer-XL time series model to generate a self-awareness embedding vector that includes the battery's hidden damage memory and future performance degradation trend; Step 2 specifically includes: Step 2.1: Each intelligent agent acquires time-series operating data of its energy storage unit during its historical life cycle, the data including at least voltage, current and temperature sequences; Step 2.2: Input the historical time series runtime data into the locally deployed Transformer-XL time series model; Step 2.3: The Transformer-XL time series model processes the historical time series running data and finally outputs the self-awareness embedding vector; In the process of calculating the self-attention score in the Transformer-XL model, a temperature control factor weighted attention mechanism is applied to calculate a bias weight based on temperature records for each historical time segment and add this bias weight to the calculation of the attention score.

[0032] The calculation formula for the temperature control factor weighted attention mechanism is as follows:

[0033] Where Q is the query vector, K is the key vector, and V is the value vector. The dimension of the key vector. This is the transpose operation of the key vector K, and softmax is the normalization function. For temperature control bias matrix, The calculation method is as follows:

[0034] in, Representing historical moments Battery temperature records, Indicates the battery's safe temperature threshold. This represents the thermal damage memory coefficient, where t is the current time step and j is the historical time step. This is the temperature control bias value of the current time step t relative to the historical time step j.

[0035] When the battery temperature severely exceeds the standard at a certain point in history This creates a large positive bias, causing the attention weights at the current moment to be highly focused on that historical segment. This means the agent can "recall" the high-temperature incident three years ago, thereby inferring the root cause of the current increase in internal resistance and generating more accurate data. vector.

[0036] Step 3: Each agent constructs a dynamic topology graph through the communication links between its base stations, broadcasts encrypted intent feature vectors, and aggregates neighbor features using the intent-aware graph network to form an environmental perception embedding vector that characterizes the resonance level of regional power supply risk. Step 3 specifically includes the following steps: Step 3.1: Each agent constructs a dynamic topology graph between base stations based on the geographical proximity and power supply topology of its base station, and dynamically maintains a list of one-hop neighbor nodes; Step 3.2: Each agent inputs its own base station operation requirements and energy storage unit status into a lightweight MLP encoder to generate an encrypted semantic intent feature vector containing both the request for help intent and the cooperation intent; Specifically, intelligent agents Observe it The input is fed into a lightweight MLP encoder to generate an intent vector. :

[0037] This vector implies semantic information (e.g.: The larger the value of the first dimension, the stronger the desire to "seek help".

[0038] Step 3.3: Broadcast the intent feature vector to neighboring nodes through the inter-base station communication link, without transmitting the original operation data; Step 3.4: Calculate the attention coefficients of the intent feature vectors of neighboring nodes using the graph attention mechanism, aggregate the neighboring features according to the coefficient weights, and generate the environment-aware embedding vector.

[0039] In step 3.4, the environment-aware embedding vector Calculated using the following formula:

[0040] in, These are the attention weight coefficients of the intention feature vector of neighboring agent j. W It is a shared weight matrix. It is the intention feature vector of neighboring agent j. It is the activation function, N( i ) is the intelligent agent i i The set of neighbors; Attention coefficient Indicates neighbors Intent against base stations Importance:

[0041] If neighbors are in the same vulnerable power supply area It issued a strong appeal for help. (large modulus) It will automatically increase in size, making Sensing environmental risks, they tend to retain their own batteries (suppressing sell orders), thus achieving group risk management.

[0042] Risk consensus.

[0043] Step 4: Each agent combines its self-awareness embedding vector and environmental perception embedding vector, and integrates its own base station operation requirements and energy storage unit status to autonomously generate a buy or sell price for asset swap, forming a virtual bidding strategy based on risk pricing. In step 4, the generation of the virtual bidding strategy follows the principle of diminishing marginal utility, specifically: When an agent predicts that the current battery health degradation will lead to a future power supply capacity gap, it generates a buy offer based on the marginal utility of the expected increase in lifespan; the growth rate of the buy offer slows down as the offer increases, and the upper limit of the offer is limited by the credit score balance in the agent's virtual account; When the agent determines that the current battery health is sufficient and the risk to itself after the sale is controllable, it generates a selling price based on the reduction in its own survival utility caused by the sale of the battery; the selling price is the lowest acceptable price that the agent is willing to accept for the sale of its battery entity.

[0044] Specifically, define the intelligent agent utility function :

[0045] That is: Utility = Survival time gain - Payment cost + Remaining credit score value.

[0046] The Actor network outputs a quote. Mathematically, the optimal offer should be close to its marginal survival value:

[0047] in This represents the risk aversion coefficient. When a base station predicts it is about to experience a downtime, the partial derivative tends towards infinity, leading to a surge in pricing. Soaring.

[0048] The cloud-based matching engine maintains an order book. For purchases... and sell orders The conditions for the transaction to be completed are: Cost constraint verification formula:

[0049] in: The unit is "virtual credit score"; The unit is "legal tender"; The shadow price exchange rate set for the system, for example, is set as follows: The "points / yuan" means that for every yuan the operations and maintenance team spends on engineering costs, the system needs to consume at least 10 virtual credit points from the price difference surplus of both parties in the transaction, thus ensuring the economic effectiveness of physical reconstruction. Based on geographical distance Logistics transportation functions.

[0050] The system is trained through multiple rounds of game theory, with the goal of minimizing the regret values ​​of all agents. When the system converges to a Nash equilibrium, it satisfies:

[0051] This means that, under the current market price and neighbor strategy, no single base station can gain more profit by unilaterally changing its price, thus achieving Pareto optimal allocation of network resources.

[0052] Step 5: Simulate multiple rounds of asset swap bidding by training a policy network with a distributed execution architecture in a centralized manner. Search for a Nash equilibrium strategy combination that maximizes the long-term survival probability of the entire network and minimizes logistics entropy. After reaching an asset swap contract that satisfies cost constraints, simultaneously execute the physical reconstruction of energy storage units, the on-path migration of digital twins, and the hot loading of model memory.

[0053] Unit physical reconfiguration: Instructs maintenance personnel to perform physical handling and exchange of energy storage battery entities; Digital twin migration on the road: The entire lifecycle historical time-series data (including historical temperature extremes and voltage / current logs) of the energy storage unit of the source base station is completely encrypted and transmitted to the edge computing unit of the target base station. Step 5, the model memory hot loading specifically includes: The target base station agent receives and decrypts data packets from the source base station; The decrypted data is loaded into its built-in Transformer-XL time series model to update the model's historical memory of the current physical battery; based on the updated memory, the battery's health status and degradation trend are continuously evaluated.

[0054] In step 5, the implementation of the centralized training distributed execution architecture specifically includes: During the offline training phase, a central critic network is constructed, which receives state information from all agents and calculates the global value function and advantage function. Using the global advantage function, the local policy networks of all agents are trained simultaneously. During the online execution phase, each agent uses only the trained local policy network to make independent decisions based on its own and its neighbors' information.

[0055] Secondly, in order to train the aforementioned complex game strategy, this invention employs the Multi-Agent Proximal Policy Optimization (MAPPO) algorithm and introduces counterfactual baselines to solve the reputation allocation problem.

[0056] Loss function: Total loss function It consists of three parts:

[0057] Advantage function and counterfactual baseline: To determine exactly how much a particular base station's bidding behavior contributes to the overall network, we calculate the dominance function. When introducing a counterfactual baseline:

[0058] Q-value represents the overall survival probability across the entire network.

[0059] Second item (baseline): Assuming an intelligent agent Take other actions Meanwhile, other intelligent agents maintain their actions. The expected Q value when it remains constant.

[0060] Through this differential calculation, if a base station “sacrifices its own interests (by selling batteries at a low price)” to greatly improve the overall network survival rate, the formula will give it a very high positive reward, thereby guiding the agent to learn a cooperative strategy that balances “altruism” and “self-interest”.

[0061] Example 2

[0062] This embodiment proposes a distributed collaborative scheduling system for energy storage resources of communication base stations, the system comprising: Distributed intelligent agent nodes are deployed at various communication base stations to perform local state perception, cognitive reasoning, and bidding decision-making. The graph communication consensus module is used to maintain dynamic neighbor relationships between base stations and to handle the broadcasting and aggregation of intent feature vectors. The virtual auction matching engine, deployed in the cloud, is used to receive and match buy and sell quotes, verify cost constraints, and generate scheduling instructions. The policy evolution training platform, deployed on a server, is used to perform centralized reinforcement learning training and update and distribute policy parameters to each agent node.

[0063] The circuits, electronic components, and modules involved are all existing technologies, which can be fully implemented by those skilled in the art, and need not be elaborated upon. The scope of protection of this invention does not involve any improvement to the software and methods.

[0064] 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.

[0065] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A dynamic reconfiguration method for energy storage arrays in communication base stations based on multi-dimensional operating condition coupling, characterized in that, include: Step 1: Model each communication base station and its associated energy storage unit in the target area as an independent reinforcement learning agent, and associate each agent with the corresponding base station operation requirements and energy storage unit status; Step 2: Each agent processes the historical lifecycle operation data of the energy storage unit based on the Transformer-XL time series model to generate a self-awareness embedding vector that includes the battery's hidden damage memory and future performance degradation trend; Step 3: Each agent constructs a dynamic topology graph through the communication links between its base stations, broadcasts encrypted intent feature vectors, and aggregates neighbor features using the intent-aware graph network to form an environmental perception embedding vector that characterizes the resonance level of regional power supply risk. Step 4: Each agent combines its self-awareness embedding vector and environmental perception embedding vector, and integrates its own base station operation requirements and energy storage unit status to autonomously generate a buy or sell price for asset swap, forming a virtual bidding strategy based on risk pricing. Step 5: Simulate multiple rounds of asset swap bidding by training a policy network with a distributed execution architecture in a centralized manner. Search for a Nash equilibrium strategy combination that maximizes the long-term survival probability of the entire network and minimizes logistics entropy. After reaching an asset swap contract that satisfies cost constraints, simultaneously execute the physical reconstruction of energy storage units, the on-path migration of digital twins, and the hot loading of model memory.

2. The method for dynamic reconfiguration of communication base station energy storage array based on multi-dimensional operating condition coupling according to claim 1, characterized in that, In step 1, each of the intelligent agents is configured with a local observation space and a virtual utility account; The local observation space is used to observe the local load of the corresponding base station, the real-time voltage of the auxiliary energy storage unit, the operating environment temperature, and the historical operation and maintenance logs of both. The virtual utility account is used to pay or receive virtual credit points during the bidding process, quantifying the urgency of the corresponding base station's demand for high-quality energy storage resources.

3. The method for dynamic reconfiguration of communication base station energy storage array based on multi-dimensional operating condition coupling according to claim 1, characterized in that, Step 2 specifically includes: Step 2.1: Each intelligent agent acquires time-series operating data of its energy storage unit during its historical life cycle, the data including at least voltage, current and temperature sequences; Step 2.2: Input the historical time series runtime data into the locally deployed Transformer-XL time series model; Step 2.3: The Transformer-XL time series model processes the historical time series running data and finally outputs the self-awareness embedding vector; In the process of calculating the self-attention score in the Transformer-XL model, a temperature control factor weighted attention mechanism is applied to calculate a bias weight based on temperature records for each historical time segment and add this bias weight to the calculation of the attention score.

4. The method for dynamic reconfiguration of a communication base station energy storage array based on multi-dimensional operating condition coupling according to claim 2, characterized in that, The calculation formula for the temperature control factor weighted attention mechanism is as follows: Where Q is the query vector, K is the key vector, and V is the value vector. The dimension of the key vector. This is the transpose operation of the key vector K, where softmax is the normalization function. For temperature control bias matrix, The calculation method is as follows: in, Representing historical moments Battery temperature records, Indicates the battery's safe temperature threshold. This represents the thermal damage memory coefficient, where t is the current time step and j is the historical time step. This is the temperature control bias value of the current time step t relative to the historical time step j.

5. The method for dynamic reconfiguration of a communication base station energy storage array based on multi-dimensional operating condition coupling according to claim 4, characterized in that, Step 3 specifically includes the following steps: Step 3.1: Each agent constructs a dynamic topology graph between base stations based on the geographical proximity and power supply topology of its base station, and dynamically maintains a list of one-hop neighbor nodes; Step 3.2: Each agent inputs its own base station operation requirements and energy storage unit status into a lightweight MLP encoder to generate an encrypted semantic intent feature vector containing both the request for help intent and the cooperation intent; Step 3.3: Broadcast the intent feature vector to neighboring nodes through the inter-base station communication link, without transmitting the original operation data; Step 3.4: Calculate the attention coefficients of the intent feature vectors of neighboring nodes using the graph attention mechanism, aggregate the neighboring features according to the coefficient weights, and generate the environment-aware embedding vector.

6. The method for dynamic reconfiguration of a communication base station energy storage array based on multi-dimensional operating condition coupling according to claim 5, characterized in that, In step 3.4, the environment-aware embedding vector Calculated using the following formula: in, These are the attention weight coefficients of the intention feature vector of neighboring agent j. W It is a shared weight matrix. It is the intention feature vector of neighboring agent j. It is the activation function, N( i ) is the intelligent agent i i The set of neighbors.

7. The method for dynamic reconfiguration of a communication base station energy storage array based on multi-dimensional operating condition coupling according to claim 1, characterized in that, In step 4, the generation of the virtual bidding strategy follows the principle of diminishing marginal utility, specifically: When an agent predicts that the current battery health degradation will lead to a future power supply capacity gap, it generates a buy offer based on the marginal utility of the expected increase in lifespan; the growth rate of the buy offer slows down as the offer increases, and the upper limit of the offer is limited by the credit score balance in the agent's virtual account; When the agent determines that the current battery health is sufficient and the risk to itself after the sale is controllable, it generates a selling price based on the reduction in its own survival utility caused by the sale of the battery; the selling price is the lowest acceptable price that the agent is willing to accept for the sale of its battery entity.

8. The method for dynamic reconfiguration of energy storage arrays for communication base stations based on multi-dimensional operating condition coupling according to claim 1, characterized in that, Step 5, specifically the model memory hot loading, includes: The target base station agent receives and decrypts data packets from the source base station; The decrypted data is loaded into its built-in Transformer-XL time series model to update the model's historical memory of the current physical battery; based on the updated memory, the battery's health status and degradation trend are continuously evaluated.

9. The method for dynamic reconfiguration of a communication base station energy storage array based on multi-dimensional operating condition coupling according to claim 1, characterized in that, The implementation of the centralized training distributed execution architecture specifically includes: During the offline training phase, a central critic network is constructed, which receives state information from all agents and calculates the global value function and advantage function. Using the global advantage function, the local policy networks of all agents are trained simultaneously. During the online execution phase, each agent uses only the trained local policy network to make independent decisions based on its own and its neighbors' information.

10. A distributed collaborative scheduling system for energy storage resources of a communication base station for implementing the method of any one of claims 1 to 9, characterized in that, The system includes: Distributed intelligent agent nodes are deployed at various communication base stations to perform local state perception, cognitive reasoning, and bidding decision-making. The graph communication consensus module is used to maintain dynamic neighbor relationships between base stations and to handle the broadcasting and aggregation of intent feature vectors. The virtual auction matching engine, deployed in the cloud, is used to receive and match buy and sell quotes, verify cost constraints, and generate scheduling instructions. The policy evolution training platform, deployed on a server, is used to perform centralized reinforcement learning training and update and distribute policy parameters to each agent node.