Base station energy saving method and device based on mobile world network model

Abstract

The invention provides a base station energy saving method and device based on a mobile world network model. The method comprises: determining a network state of a target area, the target area comprising a base station and a cell, the base station being used for providing a network service for the cell; aiming at reducing the energy consumption of the base station in the target area, determining an on-off action strategy for the network service of each cell in the target area according to the network state; analyzing the network state and the switching action strategy through a mobile world network model to obtain a predicted network state of the target area at the future moment corresponding to the switching action strategy; the mobile world network model is used for determining a predicted network state corresponding to the switching action strategy according to the network state, the knowledge graph corresponding to the target area and the switching action strategy; determining a target switch action strategy corresponding to the target area according to the predicted network state; and controlling the operation of the base station in the target area according to the target switch action strategy. The application can solve the problems in related technologies.

Description

Technical Field

[0001] The present application relates to the technical field of wireless communications and intelligent network optimization, and in particular to a base station energy-saving method and device based on a mobile world network model. Background Art

[0002] With the rapid development of fifth-generation mobile communication technology (5G) and the large-scale deployment of networks, the energy consumption of mobile communication base stations has become increasingly prominent. According to industry statistics, base stations account for over 60% of telecom operators' total energy consumption. This high energy cost not only puts pressure on operators' economic profitability but also becomes a key bottleneck in achieving green network construction and the industry's sustainable development goals.

[0003] Base stations are the core physical infrastructure of mobile communication networks, typically consisting of a communication tower, an antenna system mounted on the tower, and a cabinet beneath the tower (containing signal processing units, power modules, etc.). A cell is a specific coverage area created by base station equipment that provides wireless network services to users. A physical base station can be equipped with multiple antennas and radio frequency units facing different directions, thereby simultaneously generating and managing one or more cells.

[0004] To reduce network energy consumption, operators generally adopt an on-demand cell on / off mechanism, shutting down network services in some cells when traffic is low, thereby achieving energy conservation. Existing cell energy conservation methods are mostly based on the following technical approaches: 1. Static threshold-based methods: Fixed load or user number thresholds are set, and shutdown is triggered when the cell load or number of users falls below the corresponding threshold. This method is simple to implement, but cannot adapt to dynamic changes in traffic, which can easily lead to degraded service quality or limited energy conservation. 2. Heuristic strategy-based methods: Rules based on neighboring cell relationships, channel quality, and other factors are introduced to determine whether a cell can be shut down. This enhances system security and stability, but lacks accurate modeling of inter-cell migration traffic and user behavior, and the strategy remains crude. 3. Reinforcement learning-based methods: In recent years, some studies have attempted to apply reinforcement learning (RL) to cell energy conservation control, learning strategies through interaction with the environment. However, RL methods have the following shortcomings: policy convergence relies on a large amount of real feedback, resulting in high training costs; the lack of high-fidelity environment models leads to a "reality gap" between policy training and actual deployment; and the traffic migration process is usually simplified in modeling, such as assuming a linear correlation between the number of users and traffic, which fails to reflect the complex spatiotemporal coupling characteristics.

[0005] In summary, there are two key issues that have not been resolved in existing technologies: (1) the inability to accurately predict future network status, resulting in lagging and conservative energy-saving strategies; and (2) the lack of a high-precision assessment mechanism for post-migration load and signal quality, which can easily lead to a decline in coverage or service quality. Summary of the Invention

[0006] The present application provides a base station energy-saving method and device based on a mobile world network model, which can accurately predict future network status, accurately model user behavior, and achieve high-precision evaluation of post-migration load and signal quality, effectively solving problems in the existing technology.

[0007] This application provides a base station energy saving method based on a mobile world network model, comprising the following steps: Determining a network status of a target area, where the target area includes a base station and a cell, the base station being configured to provide network services for the cell; With the goal of reducing energy consumption of base stations in the target area, determining, based on the network status, a switching action strategy for network services of each cell in the target area; Analyzing the network state and the switching action strategy using the mobile world network model to obtain a predicted network state of the target area at a future time corresponding to the switching action strategy, wherein the mobile world network model is used to perform semantic analysis on network influencing factors of the target area based on the network state and a knowledge graph established based on target objects and relationships between target objects in the target area, and determining the predicted network state corresponding to the switching action strategy based on the semantic analysis results, the switching action strategy, and the network state; Determining a target switching action strategy corresponding to the target area according to the predicted network state; The operation of the base stations in the target area is controlled according to the target switch action strategy.

[0008] The present application also provides a base station energy-saving device based on a mobile world network model, comprising the following modules: A first determining module is configured to determine a network status of a target area, where the target area includes a base station and a cell, and the base station is configured to provide network services for the cell; a second determining module, configured to determine, based on the network status, a switching action strategy for network services of each cell in the target area with the goal of reducing energy consumption of base stations in the target area; a prediction module, configured to analyze the network state and the switching action strategy using the mobile world network model to obtain a predicted network state of the target area at a future time corresponding to the switching action strategy, wherein the mobile world network model is configured to perform semantic analysis on network influencing factors of the target area based on the network state and a knowledge graph established based on target objects and relationships between target objects in the target area, and determine the predicted network state corresponding to the switching action strategy based on the semantic analysis results, the switching action strategy, and the network state; a third determining module, configured to determine a target switching action strategy corresponding to the target area according to the predicted network state; The control module is configured to control the operation of the base stations in the target area according to the target switch action strategy.

[0009] The present application also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the base station energy saving method based on the mobile world network model as described above is implemented.

[0010] The present application also provides a non-transitory computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the base station energy saving method based on the mobile world network model as described above is implemented.

[0011] The present application also provides a computer program product, including a computer program, which, when executed by a processor, implements any of the above-described base station energy saving methods based on a mobile world network model.

[0012] The base station energy saving method and device based on the mobile world network model provided in this application have at least the following technical effects: First, the mobile world network model is used to perform forward-looking analysis of a given network state and switching action strategy. This can generate and evaluate the predicted network state of the target area at future moments after the switching action strategy is implemented. This addresses the problem that traditional methods, due to their inability to accurately predict future network states, can only make passive and reactive decisions, resulting in lagging and conservative energy-saving strategies.

[0013] Second, the knowledge graph is used to perform semantic analysis on the relationship between target objects, which can deeply consider the network influencing factors of the target area. Since the relationship between target objects can be defined according to actual needs, user behavior can be deeply modeled through the knowledge graph. When predicting the future network status of the target area, the complex internal relationship between multi-dimensional factors such as user behavior, cell load, and signal quality can be accurately understood. This solves the problem that existing methods lack a high-precision evaluation mechanism for the post-migration status and make decisions based on a single, rough indicator, which easily leads to coverage gaps or reduced service quality. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] In order to more clearly illustrate the technical solutions in the present application or the prior art, a brief introduction will be given below to the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0015] Figure 1 This is a flowchart of a base station energy saving method based on a mobile world network model shown in an embodiment of the present application.

[0016] Figure 2 This is a schematic diagram of the internal principles of a mobile world network model shown in an embodiment of the present application.

[0017] Figure 3 This is a schematic diagram of the working principle of an energy-saving decision-making system shown in an embodiment of the present application.

[0018] Figure 4 This is a schematic diagram of the energy efficiency saving rate under the strategy shown in one embodiment of the present application.

[0019] Figure 5 This is a structural block diagram of a base station energy-saving device based on a mobile world network model shown in an embodiment of the present application.

[0020] Figure 6 This is a schematic diagram of the physical structure of an electronic device shown in an embodiment of the present application. DETAILED DESCRIPTION

[0021] To make the objectives, technical solutions, and advantages of this application more clear, the technical solutions of this application will be clearly and completely described below in conjunction with the accompanying drawings. Obviously, the embodiments described are only part of the embodiments of this application, not all of them. Based on the embodiments of this application, all other embodiments obtained by ordinary technicians in this field without making any creative efforts are within the scope of protection of this application.

[0022] Most existing cell energy-saving methods are based on the following technical paths: 1. Static threshold-based approach. For example, if a cell in an office area is closed if there are fewer than 20 people, then late at night, emergency calls from users in the basement might fail due to poor backup signals. However, if the threshold is set extremely low (for example, 5 people) to avoid this problem, the cell will almost never be closed, wasting power without achieving energy savings. This fixed approach fails to dynamically balance service and energy savings.

[0023] 2. Heuristic-based approaches. For example, a policy stipulates that before shutting down Cell A (covering university dormitories), the system checks whether the total load of neighboring Cell B is below 70%. Late at night, the load on Cell A decreases, and the system finds that the load on Cell B is only 50%, meeting the conditions and shutting down Cell A. However, this policy does not accurately model user behavior. At this time, a large number of students were playing high-bandwidth, low-latency online games. Their collective migration did not distribute traffic evenly, but instead caused instantaneous, concentrated network congestion at the edge of Cell B, resulting in severe game lag and impacting the network experience of other users in Cell B.

[0024] 3. Reinforcement learning-based approaches. For example, a carrier invested significant resources in building a highly simulated urban digital twin environment to train an AI (reinforcement learning) energy-saving model. However, within this virtual environment, the AI ​​performed hundreds of millions of simulated on-off cycles to learn how to respond to various traffic patterns. This process itself consumed enormous computing resources and took a long time (high training cost). Furthermore, the AI ​​had learned a perfect strategy: the historical load in a park's cell A on weekend evenings was extremely low, allowing it to be safely shut down. However, a large-scale music festival suddenly took place that weekend in the park—an unexpected event not reflected in either the historical data or the simulation model (reality gap). Furthermore, the AI ​​shut down cell A as planned on the evening of the festival. However, its internal model simplistically assumed that traffic flow was proportional to the number of users, failing to anticipate that tens of thousands of attendees simultaneously livestreaming and sharing high-definition videos would generate traffic several orders of magnitude higher than that of ordinary users (simplified modeling). Ultimately, communication services at the festival and a large area surrounding it were completely paralyzed.

[0025] In summary, there are two key issues that remain unresolved in existing technologies: (1) The inability to accurately predict future network status results in delayed and conservative energy-saving strategies. (2) The lack of a high-precision assessment mechanism for post-migration load and signal quality can easily lead to degradation of coverage or service quality.

[0026] To address the above two problems, this application provides a mobile network energy-saving control method with forward-looking and refined modeling capabilities. By introducing a mobile network world model, it accurately estimates the future network status and user distribution, provides a reward signal that conforms to the operation of real-world mobile networks for the reinforcement learning decision module, learns effective energy-saving control strategies, and achieves a dynamic optimal balance between cell energy consumption and network performance.

[0027] This application models cell shutdown control as a constrained combinatorial optimization problem with the goal of minimizing overall network energy consumption. The constraints include cell capacity, coverage integrity, and Reference Signal Received Power (RSRP) signal strength threshold.

[0028] First, define the following variables: binary variables , continuous variables , continuous variables ,in, The index number of the cell.

[0029] in, Indicates a cell Working status (1 is on, 0 is off / sleep), ; Indicates the traffic volume of the cell. Indicates the number of users in the cell.

[0030] The optimization objective function is: Among them, the cell energy consumption model is: in, is the linear energy consumption coefficient, This is the fixed cost after the cell is opened. is the energy consumption in off / sleep mode.

[0031] Based on collaboration with mobile network operators, this application summarizes and models the constraints of real-world mobile network operation and base station collaboration. The constraints include: Capacity constraints: ; Coverage constraint: All users need to be covered by at least one active cell; RSRP constraint: The average received signal strength (RSRP) of all users must be higher than the set threshold.

[0032] The following describes in detail how this application solves the above-mentioned constrained combinatorial optimization problem from the perspective of technical implementation. Figure 1This is a flow chart of a base station energy saving method based on a mobile world network model according to an embodiment of the present application. Figure 1 The base station energy saving method based on the mobile world network model of the present application may include the following steps: Step 101: Determine the network status of a target area. The target area includes base stations and cells. The base stations are used to provide network services for the cells.

[0033] The execution subject of the method of the present application is an energy-saving decision-making system, and its deployment location can be set according to actual needs.

[0034] The target area may be any pre-set geographical area, and this application does not impose any specific limitation on the scope of the target area.

[0035] Specifically, the network status of the target area may include: the traffic of each cell (uplink / downlink traffic), the number of users in each cell (mainly the number of active users), the utilization rate of the physical resource blocks of each cell, the switch status of the network service of each cell (i.e., whether it is on or off / dormant), the geographic location of the user (mainly the geographic location of active users), timestamp, etc.

[0036] In this application, the energy-saving decision-making system can make a decision every preset time period. Each time a decision is made, step 101 is executed to collect network status data for the target area at the current moment. The timestamp in the obtained network status data is the current moment (the moment of data collection). The preset time period can be set according to actual needs.

[0037] Step 102: With the goal of reducing the energy consumption of base stations in the target area, determine a switching action strategy for network services of each cell in the target area according to the network status.

[0038] In this application, the energy-saving decision-making system is equipped with a reinforcement learning decision-making module. After determining the current network status, the reinforcement learning decision-making module uses the Multi-Agent Proximal Policy Optimization (MAPPO) algorithm and the Deep Deterministic Policy Gradient (DDPG) algorithm to preliminarily determine one or more network service on / off action strategies for each cell in the target area, with the goal of reducing the energy consumption of base stations in the target area.

[0039] The switching action refers to turning on or off the network service for a cell. For example, if the target cells include cell A, cell B, and cell C, then a switching action strategy for the target area can be {cell A: off; cell B: on; cell C: on}.

[0040] Step 103: Analyze the network status and the switching action strategy using the mobile world network model to obtain a predicted network status of the target area at a future time corresponding to the switching action strategy. The mobile world network model is used to perform semantic analysis on network influencing factors of the target area based on the network status and a knowledge graph established based on the relationships between target objects and target objects in the target area. The predicted network status corresponding to the switching action strategy is determined based on the semantic analysis results, the switching action strategy, and the network status.

[0041] In this application, the switching action strategy obtained in step 102 is a relatively crude strategy determined solely based on the current network status of the target area. Directly implementing this strategy would result in problems encountered in related technologies. Therefore, this application further analyzes all switching action strategies through steps 103 and 104, screening out the optimal switching action strategy from among them, i.e., the target switching action strategy.

[0042] In this application, a mobile world network model is provided in the energy-saving decision-making system. The mobile world network model includes a knowledge graph corresponding to the target area. This knowledge graph is pre-established based on various preset target objects and relationships between target objects in the target area. The target object type can be set based on actual needs.

[0043] In this application, the mobile world network model includes an environmental feature and data input module. This environmental feature and data input module can analyze various types of historical data of the target area to construct a knowledge graph of the target area. The historical data here mainly includes but is not limited to the following types: Type 1: Network infrastructure data. This type of data is relatively static and describes the physical and logical structure of the network. Type 1 further includes: (1) Network topology: A diagram of the physical adjacency between cells, which serves as the basis for embedding spatial structures and analyzing migration strategies; (2) Cell attributes: each cell’s unique ID, the physical base station it belongs to, etc.

[0044] (3) Base station attributes: antenna coverage direction, maximum capacity ( ), supported network standards (4G / 5G), operating frequency bands, etc.

[0045] Type 2: Network service and performance data. This type of data is dynamically changing and describes the current operating status of the network. Type 2 further includes: (1) Cell traffic (uplink / downlink traffic): The current real-time throughput of each cell.

[0046] (2) Number of connected users: the number of active users currently served by each cell.

[0047] (3) Physical resource block utilization rate of the cell: a key indicator to measure the degree of wireless channel resource occupancy.

[0048] (4) Cell switch status: whether the cell is in on or off / dormant state.

[0049] Type 3: User and environment data, further including: (1) User distribution data: The processed user spatial distribution heat map shows the geographical aggregation of users.

[0050] (2) User profile: user's app usage preferences, user's device type distribution, frequency of use of each app, user's frequently used geographical locations, time preferences, etc.

[0051] (3) Geographical semantics: the functional type of the community coverage area, such as shopping malls, office buildings, university dormitories, highways, etc.

[0052] Type 4: Time and task requirements.

[0053] (1) Temporal semantics: The current timestamp includes the day of the week, whether it is a public holiday, whether it is a weekday evening, whether it is the peak period of electricity consumption in summer, etc. These are the key inputs for learning spatiotemporal periodicity.

[0054] (2) External event information: external data accessed, such as the imported “local major event calendar” (concerts, sports events, etc.) or “severe weather warnings”.

[0055] In this application, the environmental characteristics and data input module can analyze a large amount of historical data covering the above-mentioned types one to four in the target area, extract valuable semantic data, and construct a knowledge graph of the target area based on these valuable semantic data. Valuable semantic data include but are not limited to network topology, cell attributes, base station attributes, geospatial semantics, user portraits, and time semantics. The target objects (entities) in the knowledge graph may include but are not limited to base stations, cells, users, devices, geographic grids / regions (a specific geographic block, whose attributes may include commercial, residential, industrial, green space, shopping malls, schools, hospitals, etc.), time periods (weekday morning rush hour, weekend afternoons, statutory holidays, etc.), Apps, etc., which can be set according to actual needs.

[0056] In the present application, the environmental feature and data input module also has a heterogeneous data coding (Heterogeneous Data Tokenization) function, which supports the unified coding of multi-source heterogeneous data such as grid, graph structure, and point. Using Conv2D / 3D, GNN / GCN, and Patchify technologies, multi-modal information is embedded into a unified embedding space to provide context for the generation module in the mobile world network model. Since the inference module in the mobile world network model only uses high-dimensional vectors for inference and cannot directly understand pictures, network graphs, or GPS coordinate points, it is necessary to use the heterogeneous data coding function to uniformly encode various grid, graph structure, and point data into high-dimensional vectors.

[0057] In summary, the environmental feature and data input module of the present application includes a semantic relationship module and a heterogeneous data coding module. The semantic relationship module is responsible for constructing a knowledge graph using city communication network structures (such as network topology, cell division, etc.) and user behavior records (such as App usage frequency, geographic location, time period preference, etc.). This knowledge graph at least encodes the semantic relationship between users and base stations (such as service relationship, belonging to a cell, activity frequency, etc.), providing a structured background for subsequent generation. The heterogeneous data coding module is responsible for embedding multi-modal information (grid information, graph structure information, and point information) into a unified representation space.

[0058] In the present application, the knowledge graph can be a time-series knowledge graph, i.e., the relationship between each target object can be stored in the form of a timestamp.

[0059] In step 103, the mobile world network model sends the network state at the current time to the environmental feature and data input module. The environmental feature and data input module performs semantic analysis on the network influencing factors in the target area based on the network state at the current time and the knowledge graph, obtaining a semantic analysis result. Then, the mobile world network model predicts the network state that the target area may appear in the future (predicted network state) after executing the switch action strategy on the base stations in the target area based on the network state at the current time, according to the semantic analysis result, the switch action strategy, and the network state. The future time can be one or more, and the interval between each future time can be set according to actual needs.

[0060] Step 104, determining the target switch action strategy corresponding to the target area according to the predicted network state.

[0061] When step 104 is executed, the reinforcement learning decision module determines the optimal switching action strategy corresponding to the target area, ie, the target switching action strategy, based on each predicted network state.

[0062] Step 105: Control the operation of the base stations in the target area according to the target switch action strategy.

[0063] Execute step 105, the reinforcement learning decision module converts the target switch action strategy into instructions that can be recognized and executed by the relevant physical devices in the base station that needs to be controlled, and then sends the obtained instructions to the relevant physical devices in the base station that needs to be controlled. Finally, after executing the instructions, these relevant physical devices can effectively reduce the energy consumption of the base station.

[0064] The base station energy-saving method based on the mobile world network model provided in this application can be specifically applied to the energy consumption optimization of base stations in 4G / 5G heterogeneous networks.

[0065] The method of this application has at least the following technical effects: First, the mobile world network model is used to perform forward-looking analysis of a given network state and switching action strategy. This can generate and evaluate the predicted network state of the target area at future moments after the switching action strategy is implemented. This addresses the problem that traditional methods, due to their inability to accurately predict future network states, can only make passive and reactive decisions, resulting in lagging and conservative energy-saving strategies.

[0066] Second, the knowledge graph is used to perform semantic analysis on the relationship between target objects, which can deeply consider the network influencing factors of the target area. Since the relationship between target objects can be defined according to actual needs, user behavior can be deeply modeled through the knowledge graph. When predicting the future network status of the target area, the complex internal relationship between multi-dimensional factors such as user behavior, cell load, and signal quality can be accurately understood. This solves the problem that existing methods lack a high-precision evaluation mechanism for the post-migration status and make decisions based on a single, rough indicator, which easily leads to coverage gaps or reduced service quality.

[0067] In conjunction with the above embodiments, in one implementation, there are multiple switching action strategies, each of which corresponds to a predicted network state of a target area at a future time. In this case, in step 104, determining the target switching action strategy corresponding to the target area based on the predicted network state may include: Step 1041: Determine the network energy consumption and network performance of the target area after executing each switch action strategy based on each predicted network state.

[0068] In this application, in order to enable the energy-saving strategy to reasonably evaluate its effectiveness, this application designs an environment modeling module independent of the mobile world network model in the energy-saving decision-making system. This module is mainly responsible for providing network performance and energy consumption estimates as a reward signal when the reinforcement learning decision-making module makes decisions.

[0069] Therefore, step 1041 is executed to analyze each predicted network state through the environment modeling module to determine the network energy consumption and network performance of the target area after executing each switch action strategy.

[0070] Step 1042: Determine the evaluation score corresponding to each switch action strategy based on network energy consumption and network performance.

[0071] In step 1042, the environment modeling module provides a two-dimensional reward signal by combining network energy consumption and network performance feedback, ensuring that the reinforcement learning decision module considers the balance between network energy efficiency and user experience when optimizing the strategy.

[0072] In this application, the environmental modeling module can also analyze each predicted network state to check whether it meets the capacity constraints, coverage constraints and RSRP constraints mentioned above, and give a penalty signal based on the inspection results. Finally, the evaluation score corresponding to each switching action strategy is obtained based on the reward signal and the penalty signal.

[0073] Step 1043: Determine a target switching action strategy corresponding to the target area from among the multiple switching action strategies based on the evaluation scores.

[0074] Execute step 1043 , for each evaluation score, first filter out the evaluation scores that are greater than the preset score threshold, and then determine the switch action strategy corresponding to the highest score among the filtered scores as the target switch action strategy.

[0075] In this application, multiple candidate switching action strategies are deduced, and a comprehensive evaluation score is generated based on network energy consumption and network performance. Multiple possible switching action strategies can be quantitatively compared and weighed within a decision cycle, so as to select the target switching action strategy with the highest score. This can solve the problem that traditional strategies have a single decision path, resulting in an unsatisfactory balance between energy saving and performance.

[0076] In combination with the above embodiments, in one implementation, in step 1041, determining the network energy consumption and network performance of the target area after executing each switching action strategy based on each predicted network state may include: Based on each predicted network state, determine the energy consumption of the remote radio unit and baseband unit of the base station in the target area after executing each switching action strategy, as well as the channel strength and perceived network rate of different users in the target area under the current predicted network state; Determine the network energy consumption in the target area based on the energy consumption of the remote radio unit and the baseband unit; Determine the network performance of the target area based on channel strength and network rate.

[0077] The energy consumption of the remote radio frequency units of the base stations in the target area after executing each switching action strategy according to each predicted network state may include: For each switching action strategy, the energy consumption of the remote radio unit of the base station in the target area is determined based on the traffic, number of users, fixed energy consumption after startup, and sleep energy consumption of each cell in the target area after execution.

[0078] In this application, the environment modeling module includes an energy consumption estimation module and a network performance feedback module. The energy consumption estimation module is based on the energy consumption model given above. , substitute the traffic of each cell in the predicted network status and the number of users in each cell , estimates the energy consumption of the Remote Radio Unit (RRU) to obtain the energy consumption of base stations within the target area. Furthermore, the energy consumption estimation module uses a fixed average to approximate the energy consumption of the Baseband Unit (BBU). Finally, these two energy consumption components are combined to obtain the total network energy consumption for the target area. By incorporating the energy consumption estimation module, the energy-saving decision system can accurately reflect the energy consumption changes of base stations under different loads.

[0079] In this application, in order to ensure that energy-saving measures do not lead to a decline in network performance, the network performance feedback module combines the 3GPP standard channel modeling method to simulate the channel strength of different users in the target area under the predicted network state. At the same time, the user's perceived rate is calculated through the Shannon formula as an indicator to measure the service quality.

[0080] In this application, if there are multiple predicted network states, for each predicted network state, the network energy consumption and network performance of the target area under the predicted network state need to be determined according to the above method.

[0081] This application decomposes the evaluation indicators into a refined energy consumption model based on physical components (RRU, BBU) and a physical layer performance model based on individual users (channel strength, perceived rate). When evaluating a switching action strategy, it can achieve a high-fidelity, quantifiable comprehensive consideration of the strategy's actual energy consumption and real user experience. This solves the problem that existing methods rely solely on macro and indirect indicators (such as total load rate) for evaluation, resulting in huge deviations between their estimates of actual energy consumption and predictions of real user experience, thereby making incorrect decisions.

[0082] In combination with the above embodiments, in one implementation, in step 102, determining a switching action strategy for network services of each cell in the target area according to the network status may include: Step 1021: According to the network status, the on / off status of the network service for each cell in the target area is determined through the multi-agent proximal strategy optimization algorithm.

[0083] Execute step 1021 to perform reinforcement learning using the MAPPO algorithm on the on / off state of each cell. Each cell is considered an independent intelligent agent, making sleep / activation decisions based on its local state while coordinating with the states of other cells. The entire energy-saving decision-making system uses a centralized reward mechanism to optimize the synergy between cells and ensure maximum overall network energy efficiency.

[0084] The MAPPO algorithm is used to determine which cells in a target area should have network services enabled and which should be dormant. Specifically, this method treats each cell as an independent agent. This multi-agent architecture allows each cell to make preliminary judgments based on its local state (determined by the mobile world network model). However, rather than making decisions based solely on its own circumstances, it is trained and guided by a centralized reward mechanism. Ultimately, this "individual decision-making, collective reward" approach incentivizes each cell agent to collaborate rather than make selfish decisions, thereby finding a combination of on / off strategies that maximizes energy efficiency in the target area.

[0085] For example, the target area includes cells A, B, and C. Cell A, based on its current network status and after coordinating with the current network status of other cells, reports a decision to the reinforcement learning decision module: "Cell A's network service can be turned off." Similarly, cell B reports a decision to the reinforcement learning decision module: "Cell B's network service needs to remain on." Cell C's reinforcement learning decision module reports a decision: "Cell C's network service needs to remain on." Finally, the reinforcement learning decision module obtains the on / off status of cells A, B, and C.

[0086] Step 1022: Based on the on / off status of the network services in each cell in the target area, a migration strategy is determined by using a deep deterministic policy gradient algorithm to migrate the network services of users in a cell with an off status to a cell with an on status.

[0087] Execute step 1022. For cells that need to be shut down, the DDPG algorithm is used to optimize user migration strategies. In this case, the reinforcement learning decision module can accurately predict and simulate actual user behavior based on the current network status and output fine-grained user diversion ratios to ensure balanced network load and a stable user experience.

[0088] Specifically, after the MAPPO algorithm decides to shut down a certain cell (such as cell A), it needs to solve the problem of how to divert a precise and continuous number of users. DDPG is an RL algorithm that processes continuous action spaces. In this application, its task is to output one or a set of specific and continuous diversion ratio values. For example, a precise action can be output: migrate 65.7% of users in cell A to cell B, and the remaining 34.3% of users to cell C. Through this fine-grained control, DDPG can perform the smoothest and most balanced user redistribution based on the real-time and subtle load differences in neighboring cells, avoiding the secondary congestion that may be caused by roughly migrating all users to one neighboring cell, and maximizing the user experience during the migration process.

[0089] Step 1023: Determine a switching action strategy based on the switching status and migration strategy of the network services of each cell in the target area.

[0090] Step 1023 is executed, where the reinforcement learning decision module obtains a switching action strategy based on the switching status of the network services of each cell in the target area and the obtained migration strategy.

[0091] This application sets up a reinforcement learning decision-making module in the energy-saving decision-making system, so that the entire system can make efficient energy-saving decisions in real time when facing complex network environments, and gradually improve system performance through policy optimization. In addition, this application decomposes complex mixed action decisions into two problems: macro-switch state decision-making and micro-user migration decision-making, and matches the optimal algorithm (MAPPO and DDPG) for each problem. It can achieve efficient divide-and-conquer solutions for discrete switching problems that require multi-agent collaboration and continuous diversion problems that require refined control. It can solve the problems of low learning efficiency, difficult strategy convergence, and insufficiently refined final decisions caused by high problem complexity and huge action space when directly performing reinforcement learning optimization in complex mixed action spaces.

[0092] In conjunction with the above embodiments, in one implementation, the mobile world network model includes a reasoning module, which includes multiple expert modules and a routing module. Accordingly, in step 103, the mobile world network model analyzes the network state and the switching action strategy to obtain a predicted network state of the target area at a future time corresponding to the switching action strategy, which may include: Step 1031: Analyze the semantic analysis results, switch action strategies, and network status through the routing module to determine the target expert module among the multiple expert modules, as well as the target data that each target expert module needs to process in the semantic analysis results, switch action strategies, and network status.

[0093] In this application, the inference module is a Mixture-of-Experts (MoE) transformer, which consists of multiple expert modules and uses a router mechanism to dynamically select active experts, improving the model's expressiveness and efficiency. Each layer includes layer normalization, a multi-head attention mechanism, and a scalable expert sub-network, supporting detailed modeling of complex network environments.

[0094] In step 1031, the routing module performs a preliminary analysis of the semantic analysis results, switch action strategies, and network status to determine which expert modules are suitable for processing this data. The determined expert modules are then designated as target expert modules. The routing module then distributes the data suitable for processing by each target expert module to the corresponding target expert module.

[0095] Step 1032: Analyze the corresponding target data through the target expert module to obtain the predicted network state corresponding to the switch action strategy.

[0096] Execute step 1032. After receiving the target data assigned by the routing module, each target expert module analyzes the target data. Finally, the reasoning module summarizes and analyzes the data analyzed by each target expert module to obtain the predicted network state corresponding to the switch action strategy.

[0097] This application analyzes the input information through the routing module and dynamically selects a portion of the most relevant expert modules to collaboratively process specific data. It can effectively utilize an expert model library with a huge parameter scale while maintaining low computing overhead, thereby achieving refined and high-capacity modeling of complex network environments. It can solve the problem of low reasoning efficiency and difficulty in effective deployment and operation of traditional single deep learning models when pursuing higher performance due to the expansion of model parameter scale.

[0098] In conjunction with the above embodiments, in one implementation, the network status includes cell traffic, the number of users in the cell, the user's geographic location, and a timestamp. The mobile world network model includes a semantic relationship module and a reasoning module. The mobile world network model analyzes the network status and the switching action strategy to obtain a predicted network status of the target area at a future time corresponding to the switching action strategy, which may include: Through the semantic relationship module, based on the knowledge graph, the traffic volume of each cell in the target area, the number of users in each cell, the user's geographic location and timestamp are analyzed to obtain the network topology, cell attributes, base station attributes, geospatial semantics, user profiles and temporal semantics of the target area; Through the inference module, the network topology, cell attributes, base station attributes, geospatial semantics, user profiles, temporal semantics and switching action strategies of the target area are analyzed to obtain the predicted network state corresponding to the switching action strategy. The inference module is used to use the network state as the starting point and, based on the network topology, cell attributes, base station attributes, geospatial semantics, user profiles and temporal semantics of the target area, infer the predicted network state of the target area after the switching action strategy is executed on the base stations in the target area.

[0099] In this application, after collecting the current network status of the target area, the semantic relationship module can analyze the current network status based on the previously obtained knowledge graph to obtain various semantic data such as the target area's network topology, cell attributes, base station attributes, geospatial semantics, user profiles, and temporal semantics. Next, the reasoning module uses the current network status as the starting point for reasoning and, based on the various semantic data obtained by the semantic relationship module, infers the predicted network status of the target area after executing the switching action strategy for the base stations in the target area.

[0100] This application first combines real-time observation data (such as traffic and user location) with the knowledge graph through a semantic relationship module, converts it into deep semantic information including network topology, user portraits and spatiotemporal background, and then passes it to the reasoning module for decision deduction. It can place isolated data points in a complete logical context for understanding, and achieve a leap from perceptual data to cognitive scenarios. It solves the problem that traditional models can only process surface data and cannot deeply understand the causes behind it, resulting in insufficient prediction accuracy and generalization ability, and are prone to making wrong judgments in complex or unseen scenarios.

[0101] In conjunction with the above embodiments, in one implementation, the mobile world network model includes a soft prompt module. Accordingly, in step 103, the mobile world network model analyzes the network status and the switching action strategy to obtain the predicted network status of the target area at a future time corresponding to the switching action strategy, which may include: Obtain pre-configured task requirements from the soft prompt module; In accordance with the task requirements, the network status and switching action strategy are analyzed through the mobile world network model to obtain the predicted network status of the target area at the future moment corresponding to the switching action strategy.

[0102] In this application, users of the energy-saving decision-making system can pre-configure some matters that the reasoning module needs to pay attention to when performing reasoning through the soft prompt module, such as the method of determining future moments, etc., thereby improving the energy-saving decision-making effect.

[0103] Figure 2 This is a schematic diagram of the internal principle of a mobile world network model shown in an embodiment of the present application. Figure 2 , and gives a detailed introduction to each module in the mobile world network model.

[0104] exist Figure 2 In the [1], the environmental features and data input module performs the input of environmental features and mobile training data. In the communication environment encoding phase, communication environment data and user device behavior data need to be encoded. Semantic analysis is then performed on the encoded data to add valuable semantic data (such as proximity relationships, services, check-ins, device relationships, and device models) to the mobile network knowledge graph. Figure 2 Only some of the data are listed here. In actual implementation, you can choose the data type according to your needs. The multidimensional network data word representation supports the processing of raster, graph structure and point data.

[0105] The denoising network includes a feature alignment unit, a MoE network, a data-driven soft hint module, and a diffusion noise and temporal encoding module. The MoE network is the inference module mentioned above, and the data-driven soft hint module is the soft hint module mentioned above.

[0106] The feature alignment unit uses contrastive learning to align features in heterogeneous environments and introduces a scale migration mechanism for conditional adaptive modeling, making the model applicable to different regions, time periods, and load scenarios. Specifically, the feature alignment unit is used to: 1. Through the contrastive learning mechanism, achieve semantic alignment of heterogeneous features from different sources in a unified representation space to ensure the consistency of input information; 2. Through the adaptive conditional mechanism, enable features to be dynamically adjusted according to different time, space, or load scenarios, infusing the features with situational awareness. The feature alignment unit can provide semantically consistent and dynamic contextual, high-quality feature inputs to subsequent reasoning modules, improving the modeling accuracy and generalization capabilities of the mobile world network model.

[0107] The functions of each layer of the MoE network are as follows: The Mixture-of-Experts layer is used to expand the model's overall capacity in a computationally efficient manner. It comprises a router module and expert modules. The router module is a gating network that dynamically and sparsely selects and activates a small number of the most relevant expert modules based on input features to ensure computational efficiency. Each expert is an independent neural network that gradually specializes during training to focus on processing specific types of data patterns or features. The collection of all experts constitutes the model's powerful knowledge capacity.

[0108] The multi-head self attention mechanism is responsible for capturing the contextual dependencies within the input data sequence. By calculating the correlation weights between different elements in the sequence, the model can integrate global information and understand the mutual influence between different features, providing a contextual basis for accurate reasoning.

[0109] Adaptive layer normalization and residual connection (Adaptive Layernorm + Add) are used to ensure stable training of deep networks and effective flow of information.

[0110] Layer normalization is used to stabilize the input distribution of each neural network layer, preventing gradient explosion or vanishing during training and accelerating model convergence. Its adaptive nature enables it to incorporate external condition information and dynamically adjust the normalization process.

[0111] Residual connections (Add) create information shortcuts, adding the module's input directly to its output, effectively alleviating the gradient vanishing problem in deep networks, thereby supporting the construction of deeper and more complex models.

[0112] The data-driven soft prompt module is responsible for constructing query vectors for time and frequency information guidance, forming soft prompts by matching with the memory pool, and realizing multi-scale information perception and scheduling control prompt generation.

[0113] Specifically, this module first encodes the time-frequency requirements into a query vector (query), then matches them against a learnable memory pool to retrieve the most relevant knowledge vector (key). Finally, a soft prompt (prompt) is generated based on this knowledge vector and injected into the subsequent model. This mechanism enables the Mobile World Network model to dynamically adjust its internal operating mode based on different task requirements (such as minute-level or hour-level predictions), achieving controllable multi-scale information perception and generation.

[0114] The Mobile World Network model also features noise perturbing simulation. Using a diffusion model training framework, it adds noise perturbations to the input data and combines a noise scheduler with position encoding to model the denoising process, thereby achieving a stable and controllable data generation process.

[0115] Specifically, the diffusion noising and temporal encoding module describes a powerful generative model training framework. Its core function is to learn a reverse denoising process, enabling the model to grasp the inherent distribution and structure of real data, thereby enabling high-fidelity data generation. This framework first uses a noise scheduler to controllably and step-by-step denoising of real input data. Then, when the core training model receives noisy data, it accurately predicts the added noise based on the embedding of the current denoising step and the position encoding of the data. Once trained, the model iteratively applies the learned denoising steps, starting with pure random noise, to ultimately generate stable, controllable, and fresh network state data.

[0116] The data output module supports infrastructure data output and flexible scale output. To support simulation tasks for base station energy optimization, the Mobile World Network model can generate controllable simulation metrics such as the number of users per cell per time period and the uplink and downlink traffic of each base station or cell.

[0117] Specifically, the data output module generates critical infrastructure data, primarily including precise forecasts of traffic usage and user access for each cell or base station at future times. It also features flexible multi-scale modeling capabilities, supporting the generation of these metrics at varying spatial granularities (from device level to cell / grid level) and temporal granularity (from minute to hour). This multi-scale output ensures that the Mobile World Network Model can adapt to and serve energy-saving decision-making needs at different levels, providing direct and usable data support for subsequent policy evaluation and optimization.

[0118] Through the collaborative modeling of these multiple modules, the Mobile World Network Model enables comprehensive world modeling capabilities, from underlying environmental perception and historical behavior memory to task adaptation and decision support. Compared to traditional rule-based or single-task optimization approaches, it boasts stronger data adaptability, spatial generalization, and decision support capabilities. It can simulate and evaluate system responses under arbitrary configurations without interfering with the real network, providing a high-fidelity simulation environment and performance feedback for downstream tasks such as mobile network energy consumption control and parameter optimization.

[0119] The core of this application is multi-scale joint load and user modeling based on the Mobile World Network Model. By simulating the evolution of network state, the Mobile World Network Model helps the reinforcement learning decision module fully understand current and future network conditions when making base station energy-saving decisions. Within this framework, the Mobile World Network Model achieves accurate modeling of network state through the following key mechanisms: (1) Multi-scale modeling: Jointly modeling the cell-level traffic load and user-level spatial distribution to capture the spatiotemporal evolution patterns at different scales. This process enables the mobile world network model to not only predict the load changes of a single cell, but also to simultaneously consider the spatial correlation and evolution trend of user distribution.

[0120] (2) Spatiotemporal prediction: By using historical network data, the mobile world network model can learn various periodic characteristics (such as day and night traffic fluctuations, weekend effects, etc.), providing accurate time series predictions for energy-saving decisions.

[0121] (3) Spatial structure embedding: When constructing the mobile world network model, the spatial relationship between cells is extracted by constructing an adjacency graph between cells, which further helps the model capture the synergy between adjacent cells.

[0122] Through the above mechanism, the mobile world network model can generate high-fidelity predictions of network status and provide powerful state transition simulation support for subsequent decision modules.

[0123] In this application, the mobile world network model, environment modeling module and reinforcement learning decision module exist independently, but they work closely together to support the formulation and optimization of energy-saving strategies. Figure 3 As shown in the figure, the Mobile World Network Model predicts and simulates network status, providing a virtual environment for the reinforcement learning decision-making module to understand future network changes. The environmental modeling module provides accurate energy consumption and performance feedback for each decision, ensuring the effectiveness and feasibility of energy-saving strategies. The reinforcement learning decision-making module, trained in the virtual environment of the Mobile World Network Model, optimizes each cell's switching decisions and user migration strategies, ultimately achieving the optimal balance between global energy conservation and performance. Through this multi-module collaborative approach, the energy-saving decision-making system accurately grasps network status changes in real-time decision-making and effectively reduces base station energy consumption, while ensuring uncompromising user experience.

[0124] Figure 3 This is a schematic diagram of the working principle of an energy-saving decision-making system shown in an embodiment of the present application. Figure 3 ,The core implementation process of the energy-saving decision-making system is: Step 1: Observation: The reinforcement learning decision module (RL Agent) observes the current network status in the virtual environment provided by the mobile world network model (World Model).

[0125] Step 2: Decision-making: The reinforcement learning decision module outputs a hypothetical energy-saving action (Action) based on the current state through its internal policy (Policy π).

[0126] Step 3: Prediction: The mobile world network model receives the action, performs high-fidelity inference, and outputs the predicted network state at the next moment after executing the action (Prediction of next state).

[0127] Step 4: Evaluation: The environment modeling module evaluates the predicted network state, calculates a comprehensive reward score, and feeds it back to the reinforcement learning decision module.

[0128] Step 5: Learning. The reinforcement learning decision module updates its decision-making strategy based on the reward signal it receives, enabling it to make decisions that achieve higher scores in the future.

[0129] The above steps 1 to 5 are iterated repeatedly in the virtual environment until the strategy of the reinforcement learning decision module converges to the optimal one.

[0130] To validate a mobile network energy-saving method based on the Mobile World Network model, this application builds a training and validation process based on large-scale historical data from real-world communication networks. The entire method framework consists of three main phases: pre-training the Mobile World Network model, generating controllable inferences based on the Mobile World Network model, and policy optimization training of the reinforcement learning decision module. These phases work together to achieve efficient and energy-efficient network state decision-making and simulation.

[0131] First: Pre-training of the Mobile World Network Model The mobile world network model can be trained through the following steps: obtaining network state sample data; using the reconstruction error of the network state sample data and the prediction error of the predicted network state corresponding to the network state sample data as the joint optimization objectives, and training the initial mobile world network model based on the principle of the diffusion model to obtain the mobile world network model.

[0132] Specifically, in the first stage, this application systematically preprocesses historical data collected from real communication networks, including data cleaning, feature regularization, and time alignment, to construct high-quality spatiotemporal state sequences for model learning. Based on this dataset, this application employs a diffusion model training framework to perform comprehensive unsupervised or semi-supervised pre-training of the initial mobile world network model. The core concept of this framework is to learn the intrinsic distribution of the data through a noise addition and denoising process. During training, the system's joint optimization goal is to accurately reconstruct the original network state from varying degrees of noise perturbation. This process not only optimizes the reconstruction error but also, by using the historical state sequence as a guiding condition, allows the model to learn to generate coherent future network states, thereby jointly optimizing the future state prediction error. The pre-training process covers multiple time scales (e.g., peak and trough periods) and typical spatial distribution scenarios. This powerful generative modeling capability enables the model to accurately capture the spatiotemporal evolution of key variables such as network load, user distribution, traffic patterns, and base station status. After the training of this stage is completed, the mobile world network model can provide stable and consistent environment dynamic modeling capabilities, laying a solid foundation for strategy generation in subsequent stages.

[0133] Among them, the final total loss function used when training the Mobile World Network model is as follows: For the reconstruction loss, this loss is used to force the encoder and decoder to express the current network state losslessly and efficiently. To predict the loss, this loss is used to train the state transition model so that it can accurately predict the evolution of the network state in the latent space. is a hyperparameter used to balance the importance of understanding the present and predicting the future. During the training process, all parameters of the entire mobile world network model will be optimized through gradient descent and other optimization algorithms towards a common minimum. This goal is constantly being adjusted.

[0134] 2. Controllable Inference Generation Based on the Mobile World Network Model After the pre-training of the mobile world network model is completed, the application further utilizes the mobile world network model to carry out a controllable inference task, which is used to simulate and predict the network state under different energy-saving strategies. Through the system disturbance of the input strategy variable, the model can generate future network evolution paths under multiple conditions, and predict the change trend of key indicators including the load of the cell, user distribution, connection quality, etc. This controllable generation capability enables the application to evaluate the potential impact of the strategy on network performance and energy efficiency without actually deploying the strategy, thereby realizing the rapid screening and reverse deduction analysis of the energy-saving scheme. This stage can significantly improve the data efficiency and evaluation accuracy of the system in the strategy exploration process.

[0135] 3. Strategy optimization of reinforcement learning decision module In the high-fidelity environment provided by the mobile world network model, the application designs and trains a reinforcement learning decision module for automatic optimization of cell energy-saving strategies. Specifically, the reinforcement learning decision module uses the MAPPO algorithm to jointly learn the on / off state of the cell, with the goal of maximizing the energy efficiency index of the entire network while ensuring network service quality. Each cell agent obtains prediction information of future network states through the mobile world network model, and formulates a forward-looking sleep / wake-up strategy accordingly. Meanwhile, in order to further improve the local user experience, the application introduces the DDPG algorithm to optimize the dynamic migration strategy of users among multiple cells. DDPG can model complex nonlinear migration behavior, ensuring connection stability while guiding users to gather in low-load, high-efficiency areas, thereby improving overall network operation efficiency.

[0136] Through experimental verification, the optimization results of the application are compared with existing energy-saving strategies, and the specific performance indicators are shown in Table 1. The existing threshold-based method is widely used in commercial networks, but it lacks adaptability and performance constraints, resulting in a significant decline in daytime performance, as well as a significant decline in RSRP and overall coverage quality. The heuristic method relies on greedy rule-based algorithms that iteratively shut down cells according to a fixed priority order until the coverage or capacity constraints are violated. Although this method has a certain degree of adaptability, it lacks optimization depth and is highly sensitive to parameter adjustment. The reinforcement learning-based method applies advanced multi-agent reinforcement learning (MARL) techniques to solve the cell energy-saving problem. However, these methods have significant limitations. First, MARL agents rely on past data, which may lag behind real-time dynamics, resulting in reactive control rather than forward-looking decision-making. Second, these methods usually simplify the traffic migration mechanism, assuming that traffic load and user migration ratio are linearly related, i.e. This simplification fails to capture the actual traffic usage patterns of users after migration, which depend on factors such as channel quality and base station resource availability. As a result, the reward value deviates from the actual results, introducing systematic bias and reducing the effectiveness of the policy.

[0137] Table 1 In contrast, our approach significantly reduces traffic loss and RSRP degradation, both of which are less than half of the RL baseline approach. Specifically, our mobile network world model approach outperforms existing reinforcement learning approaches in terms of traffic loss and RSRP degradation, demonstrating significant optimization effects.

[0138] Figure 4 This is a schematic diagram of the energy efficiency saving rate under the strategy shown in one embodiment of the present application. Figure 4 The energy saving rate (ESR) under different strategies is demonstrated in the paper. The mobile network world model-based method of the present application outperforms all baseline methods during the daytime and achieves the best energy efficiency saving performance in both 4G and 5G networks. Specifically, in 4G networks, the method of the present application improves the energy efficiency saving rate by 17.9% compared to the existing state-of-the-art RL method (the average ESR of the RL method is 12.373, and the mobile network world model method is 14.588), and in 5G networks, it improves by 39.1% (the average ESR of the RL method is 4.786, and the average ESR of the mobile network world model method is 6.655). These improvements are attributed to the accurate traffic prediction and refined user diversion simulation integrated by the mobile network world model, which supports performance-based reward estimation.

[0139] In summary, the method of this application achieves an excellent balance between energy efficiency and service quality. Combined with the mobile world network model framework, it greatly enhances the capabilities of the reinforcement learning decision module, thereby enabling more efficient and stable energy-saving optimization in actual mobile networks.

[0140] The following describes the base station energy-saving device based on the mobile world network model provided in this application. The base station energy-saving device based on the mobile world network model described below and the base station energy-saving method based on the mobile world network model described above can refer to each other. Figure 5 This is a structural block diagram of a base station energy-saving device based on a mobile world network model shown in an embodiment of the present application.

[0141] Reference Figure 5 The base station energy-saving device based on the mobile world network model of the present application may include: A first determining module 501 is configured to determine a network status of a target area, where the target area includes a base station and a cell, and the base station is configured to provide network services for the cell; A second determining module 502 is configured to determine, based on the network status, a switching action strategy for network services of each cell in the target area with the goal of reducing energy consumption of base stations in the target area; Prediction module 503 is configured to analyze the network state and the switching action strategy using the mobile world network model to obtain a predicted network state of the target area at a future time corresponding to the switching action strategy. The mobile world network model is configured to perform semantic analysis on network influencing factors of the target area based on the network state and a knowledge graph established based on target objects and relationships between target objects in the target area, and determine the predicted network state corresponding to the switching action strategy based on the semantic analysis results, the switching action strategy, and the network state. A third determining module 504 is configured to determine a target switching action strategy corresponding to the target area according to the predicted network state; The control module 505 is configured to control the operation of the base stations in the target area according to the target switch action strategy.

[0142] According to the base station energy-saving device 500 based on the mobile world network model of the present application, there are multiple switching action strategies, each of which corresponds to a predicted network state of the target area at a future time; the prediction module 503 includes: A first determination submodule is configured to determine, based on each of the predicted network states, the network energy consumption and network performance of the target area after executing each of the switch action strategies; A second determining submodule is configured to determine an evaluation score corresponding to each of the switch action strategies according to the network energy consumption and the network performance; The third determination submodule is configured to determine, according to each of the evaluation scores, a target switching action strategy corresponding to the target area from among the plurality of switching action strategies.

[0143] According to the base station energy saving device 500 based on the mobile world network model of the present application, the third determining module 504 includes: A fourth determination submodule is configured to determine, based on the network status, a switching status of network services for each cell in the target area by using a multi-agent proximal strategy optimization algorithm; a fifth determination submodule, configured to determine, based on the on / off status of the network services of each cell in the target area, a migration strategy for migrating the network services of users in a cell whose switch state is off to a cell whose switch state is on, using a deep deterministic policy gradient algorithm; The sixth determining submodule is configured to determine the switching action strategy according to the switching status of the network services of each cell in the target area and the migration strategy.

[0144] According to the base station energy-saving device 500 based on the mobile world network model of the present application, the mobile world network model includes a reasoning module, which includes multiple expert modules and a routing module; the prediction module 503 includes: A first analysis submodule is configured to analyze the semantic analysis result, the switch action strategy, and the network status through the routing module, and determine a target expert module among the multiple expert modules, and target data that each target expert module needs to process based on the semantic analysis result, the switch action strategy, and the network status; The second analysis submodule is configured to analyze the corresponding target data through the target expert module to obtain a predicted network state corresponding to the switch action strategy.

[0145] According to the base station energy-saving device 500 based on the mobile world network model of the present application, the network status includes the traffic volume of the cell, the number of users in the cell, the geographical location of the users, and the timestamp. The mobile world network model includes a semantic relationship module and a reasoning module. The prediction module 503 includes: A third analysis submodule is configured to analyze the traffic volume of each cell in the target area, the number of users in each cell, the geographic location of the users, and the timestamp based on the knowledge graph through the semantic relationship module to obtain the network topology, cell attributes, base station attributes, geospatial semantics, user profiles, and temporal semantics of the target area; The fourth analysis submodule is used to analyze the network topology, cell attributes, base station attributes, geospatial semantics, user portraits, time semantics and the switching action strategy of the target area through the reasoning module to obtain the predicted network state corresponding to the switching action strategy. The reasoning module is used to take the network state as the starting point and, based on the network topology, cell attributes, base station attributes, geospatial semantics, user portraits and time semantics of the target area, infer the predicted network state of the target area after the switching action strategy is executed on the base stations in the target area.

[0146] According to the base station energy-saving device 500 based on the mobile world network model of the present application, the mobile world network model includes a soft prompt module, and the prediction module 503 includes: An acquisition submodule, configured to acquire pre-configured task requirements from the soft prompt module; The fifth analysis submodule is used to analyze the network status and the switching action strategy through the mobile world network model in accordance with the task requirements, and obtain the predicted network status of the target area at a future moment corresponding to the switching action strategy.

[0147] According to the base station energy-saving device 500 based on the mobile world network model of the present application, the first determining submodule includes: a seventh determination submodule, configured to determine, based on each of the predicted network states, the energy consumption of the remote radio frequency unit and the baseband unit of the base station in the target area after executing each of the switching action strategies, as well as the channel strength and perceived network rate of different users in the target area under the current predicted network state; an eighth determining submodule, configured to determine the network energy consumption of the target area according to the energy consumption of the remote radio frequency unit and the energy consumption of the baseband unit; A ninth determination submodule is configured to determine the network performance of the target area according to the channel strength and the network rate.

[0148] According to the base station energy-saving device 500 based on the mobile world network model of the present application, the predicted network status includes the traffic volume of the cell and the number of users in the cell; the seventh determination submodule includes: The tenth determination submodule is used to determine the energy consumption of the remote radio frequency unit of the base station in the target area according to the traffic, number of users, fixed energy consumption after startup and sleep energy consumption of each cell in the target area after execution of each switching action strategy.

[0149] According to the base station energy-saving device 500 based on the mobile world network model of the present application, the mobile world network model is trained by the following steps: Get network status sample data; The reconstruction error of the network state sample data and the prediction error of the predicted network state corresponding to the network state sample data are taken as joint optimization targets, and the initial mobile world network model is trained through the principle of the diffusion model to obtain the mobile world network model.

[0150] Figure 6 This is a schematic diagram of the physical structure of an electronic device shown in an embodiment of this application. Figure 6As shown, the electronic device may include: a processor 610, a communications interface 620, a memory 630, and a communication bus 640. The processor 610, the communications interface 620, and the memory 630 communicate with each other via the communication bus 640. The processor 610 may invoke logic instructions in the memory 630 to execute a base station energy saving method based on a mobile world network model.

[0151] In addition, the logical instructions in the above-mentioned memory 630 can be implemented in the form of a software functional unit and can be stored in a computer-readable storage medium when sold or used as an independent product. Based on this understanding, the technical solution of the present application, or the part that contributes to the existing technology, or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for enabling a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk, and other media that can store program code.

[0152] On the other hand, the present application also provides a computer program product, which includes a computer program. The computer program can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute a base station energy saving method based on a mobile world network model provided by the above methods.

[0153] On the other hand, the present application also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, is implemented to execute a base station energy saving method based on a mobile world network model provided by the above methods.

[0154] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, i.e., they may be located in one location or distributed across multiple network units. Some or all of the modules may be selected based on actual needs to achieve the objectives of the present embodiment. Persons of ordinary skill in the art will be able to understand and implement the present invention without inventive effort.

[0155] Through the above description of the embodiments, those skilled in the art will clearly understand that each embodiment can be implemented using software plus a necessary general-purpose hardware platform, or of course, hardware. Based on this understanding, the essence of the above technical solution, or the portion that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a magnetic disk, or an optical disk, and includes a number of instructions for causing a computer device (such as a personal computer, server, or network device) to execute the methods described in each embodiment or certain portions of the embodiments.

[0156] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A base station energy saving method based on a mobile world network model, characterized in that: include: Determining a network status of a target area, where the target area includes a base station and a cell, the base station being configured to provide network services for the cell; With the goal of reducing energy consumption of base stations in the target area, determining, based on the network status, a switching action strategy for network services of each cell in the target area; Analyzing the network state and the switching action strategy using the mobile world network model to obtain a predicted network state of the target area at a future time corresponding to the switching action strategy, wherein the mobile world network model is used to perform semantic analysis on network influencing factors of the target area based on the network state and a knowledge graph established based on target objects and relationships between target objects in the target area, and determining the predicted network state corresponding to the switching action strategy based on the semantic analysis results, the switching action strategy, and the network state; Determining a target switching action strategy corresponding to the target area according to the predicted network state; The operation of the base stations in the target area is controlled according to the target switch action strategy.

2. The base station energy saving method based on the mobile world network model according to claim 1, characterized in that: There are multiple switching action strategies, each of which corresponds to a predicted network state of the target area at a future time; determining a target switching action strategy corresponding to the target area based on the predicted network state includes: determining, based on each of the predicted network states, network energy consumption and network performance of the target area after executing each of the switching action strategies; Determining an evaluation score corresponding to each of the switch action strategies according to the network energy consumption and the network performance; According to each of the evaluation scores, a target switching action strategy corresponding to the target area is determined from the plurality of switching action strategies.

3. The base station energy saving method based on the mobile world network model according to claim 1, characterized in that: The determining, according to the network status, a switching action strategy for the network service of each cell in the target area includes: Determining, based on the network status, the on / off status of network services for each cell in the target area through a multi-agent proximal strategy optimization algorithm; Determining, based on the on / off status of network services in each cell within the target area, a migration strategy for migrating network services of users in a cell with an off status to a cell with an on status, using a deep deterministic policy gradient algorithm; The switching action strategy is determined according to the switching status of the network services of each cell in the target area and the migration strategy.

4. The base station energy saving method based on the mobile world network model according to claim 1, characterized in that: The mobile world network model includes a reasoning module, which includes multiple expert modules and a routing module. The mobile world network model is used to analyze the network state and the switching action strategy to obtain a predicted network state of the target area at a future time corresponding to the switching action strategy, including: Analyzing the semantic analysis results, the switch action strategy, and the network status through the routing module to determine a target expert module among the multiple expert modules, and target data that each target expert module needs to process based on the semantic analysis results, the switch action strategy, and the network status; The target expert module analyzes the corresponding target data to obtain the predicted network state corresponding to the switch action strategy.

5. The base station energy saving method based on the mobile world network model according to claim 1, characterized in that: The network status includes cell traffic, the number of users in the cell, the geographic location of the users, and a timestamp. The mobile world network model includes a semantic relationship module and a reasoning module. The network status and the switch action strategy are analyzed using the mobile world network model to obtain a predicted network status of the target area at a future time corresponding to the switch action strategy, including: Through the semantic relationship module, based on the knowledge graph, the traffic of each cell in the target area, the number of users in each cell, the geographical location of users, and the timestamp are analyzed to obtain the network topology, cell attributes, base station attributes, geospatial semantics, user profiles, and temporal semantics of the target area; Through the inference module, the network topology, cell attributes, base station attributes, geospatial semantics, user profile, time semantics and the switch action strategy of the target area are analyzed to obtain the predicted network state corresponding to the switch action strategy. The inference module is used to take the network state as the starting point and, based on the network topology, cell attributes, base station attributes, geospatial semantics, user profile and time semantics of the target area, infer the predicted network state of the target area after the switch action strategy is executed on the base stations in the target area.

6. The base station energy saving method based on the mobile world network model according to claim 1, characterized in that: The mobile world network model includes a soft prompt module, and the network state and the switch action strategy are analyzed by the mobile world network model to obtain a predicted network state of the target area at a future time corresponding to the switch action strategy, including: Obtaining pre-configured task requirements from the soft prompt module; In compliance with the task requirements, the network status and the switching action strategy are analyzed through the mobile world network model to obtain a predicted network status of the target area at a future moment corresponding to the switching action strategy.

7. The base station energy saving method based on the mobile world network model according to claim 2, characterized in that: The determining, based on each of the predicted network states, the network energy consumption and network performance of the target area after executing each of the switch action strategies, includes: Determining, based on each of the predicted network states, energy consumption of a remote radio unit and a baseband unit of a base station in the target area after executing each of the switching action strategies, and channel strength and perceived network rate of different users in the target area under the current predicted network state; Determining the network energy consumption of the target area according to the energy consumption of the remote radio frequency unit and the energy consumption of the baseband unit; The network performance of the target area is determined according to the channel strength and the network rate.

8. The base station energy saving method based on the mobile world network model according to claim 7, characterized in that: The predicted network state includes the traffic volume of a cell and the number of users in the cell; and determining, based on each of the predicted network states, the energy consumption of the remote radio frequency units of the base stations in the target area after executing each of the switching action strategies, includes: For each of the switching action strategies, the energy consumption of the remote radio frequency units of the base stations in the target area is determined according to the traffic, number of users, fixed energy consumption after startup, and sleep energy consumption of each cell in the target area after execution.

9. The base station energy saving method based on the mobile world network model according to any one of claims 1 to 8, characterized in that: The mobile world network model is trained through the following steps: Get network status sample data; The reconstruction error of the network state sample data and the prediction error of the predicted network state corresponding to the network state sample data are taken as joint optimization targets, and the initial mobile world network model is trained through the principle of the diffusion model to obtain the mobile world network model.

10. A base station energy-saving device based on a mobile world network model, characterized in that: include: A first determining module is configured to determine a network status of a target area, where the target area includes a base station and a cell, and the base station is configured to provide network services for the cell; a second determining module, configured to determine, based on the network status, a switching action strategy for network services of each cell in the target area with the goal of reducing energy consumption of base stations in the target area; a prediction module, configured to analyze the network state and the switching action strategy using the mobile world network model to obtain a predicted network state of the target area at a future time corresponding to the switching action strategy, wherein the mobile world network model is configured to perform semantic analysis on network influencing factors of the target area based on the network state and a knowledge graph established based on target objects and relationships between target objects in the target area, and determine the predicted network state corresponding to the switching action strategy based on the semantic analysis results, the switching action strategy, and the network state; a third determining module, configured to determine a target switching action strategy corresponding to the target area according to the predicted network state; The control module is configured to control the operation of the base stations in the target area according to the target switch action strategy.

Images