Space-time prediction based flexible virtual cell multi-base station cooperative communication system and method

Abstract

The application discloses a kind of based on spatiotemporal prediction's elastic virtual cell multi-base station cooperative communication system and implementation method.The system uses the five-layer closed-loop architecture of global data perception and graph construction layer, spatiotemporal graph prediction layer, elastic virtual cell networking layer, prediction enhancement cooperative communication layer and closed-loop feedback optimization layer is constituted.In virtual cell, system utilizes prediction result to guide the cooperative scheduling of multi-base station resource, reduces scheduling decision range, improves scheduling efficiency and real-time nature.Meanwhile, system continuously monitors network actual operating state, and according to operation feedback, prediction model and scheduling strategy are adjusted, to maintain system stable operation.The application realizes the prospective scheduling and fine utilization of network resource, improves system capacity and resource utilization efficiency, reduces network energy consumption, improves communication service quality, is applicable to high-density user aggregation and burst traffic variation scene, and has good engineering realizable and network evolution adaptation ability.

Description

Technical Field

[0001] This invention belongs to the field of intelligent control and resource optimization technology of next-generation wireless communication networks, specifically a flexible virtual cell multi-base station cooperative communication system and implementation method based on spatiotemporal prediction.

[0002] This invention is applicable to wireless access networks in high-density, high-dynamic traffic scenarios, and is especially suitable for distributed deployment environments using open wireless access network architectures. By introducing spatiotemporal graph neural networks, deep reinforcement learning and digital twin technology, it enables predictive construction, elastic adjustment and energy-efficiency-aware scheduling of virtual cells during multi-base station collaborative communication. Background Technology

[0003] With the large-scale commercialization of fifth-generation mobile communication networks and the proposed vision for sixth-generation mobile communication networks, mobile communication networks are continuously evolving towards ultra-high capacity, ultra-low latency, and high energy efficiency. In typical scenarios such as urban core business districts, transportation hubs, and large venues, user service demands are highly concentrated, and network traffic exhibits significant spatiotemporal non-uniformity and dynamic variation characteristics, placing higher demands on the real-time scheduling and coordination capabilities of wireless access networks.

[0004] Traditional cellular communication systems typically plan and allocate resources based on fixed physical cell boundaries, resulting in a static cell structure and scheduling strategy. When user distribution or service load changes rapidly, hotspot base stations can become overloaded while adjacent base stations remain idle. This not only degrades the quality of user communication services but also leads to low overall utilization efficiency of spectrum and energy resources. This type of static network architecture is no longer sufficient to meet the operational requirements of high-density, highly dynamic service scenarios.

[0005] To address the aforementioned issues, academia and industry have proposed virtual cell or cellular-free communication architectures. These architectures break down fixed physical cell boundaries to achieve user-centric, multi-base station collaborative coverage. Simultaneously, artificial intelligence technology has been introduced into the field of wireless network control to assist in traffic prediction and resource scheduling decisions. Existing solutions typically employ neural network-based models to predict base station traffic or service load, triggering virtual cell networking or resource adjustment operations accordingly, which improves the network's responsiveness to traffic changes to some extent.

[0006] However, practical research has revealed that existing AI-based communication control schemes still suffer from the following significant technical shortcomings:

[0007] (1) Lack of effective modeling capability for network spatial topology relationships. Existing traffic prediction methods mostly treat individual base stations as independent time-series objects, or only use regularized geographical proximity relationships for modeling, failing to explicitly characterize the complex topological relationships between base stations formed by factors such as interference coupling and user migration. Since wireless access networks are essentially strongly coupled systems, ignoring the spatial correlation between base stations can easily lead to inaccurate predictions of hotspot areas during migration between base stations, especially with a significant decrease in prediction performance under complex interference environments.

[0008] (2) Lack of mechanisms for perceiving and utilizing prediction uncertainty. Existing virtual cell networking strategies are usually based on a single estimate of the prediction result, combined with fixed load or distance thresholds for decision-making. However, artificial intelligence models inevitably have prediction errors when faced with sudden events or changes in service distribution. When the prediction confidence is low, if a rigid and precise networking strategy is still adopted, it is easy to cause problems such as insufficient coverage and handover failure. Existing technologies generally lack mechanisms for dynamically and elastically adjusting virtual cell boundaries based on prediction uncertainty.

[0009] (3) The computational complexity of multi-base station collaborative scheduling is high, making it difficult to meet real-time requirements. In collaborative communication scenarios involving multiple base stations, multiple users, and multiple carriers, resource scheduling involves multi-dimensional joint optimization, which is a large-scale and complex problem, belonging to a combinatorial optimization problem that is difficult to solve. Traditional reinforcement learning methods generally suffer from slow convergence speed and lagging policy updates in large-scale action spaces, making it difficult to achieve real-time scheduling in environments with rapidly changing channel states, thus limiting their engineering application in next-generation wireless communication networks.

[0010] (4) The energy efficiency optimization methods are relatively passive and it is difficult to achieve forward-looking energy-saving control. Existing solutions rely on post-hoc hibernation or simple load judgment for energy consumption optimization. They lack a mechanism for forward-looking scheduling and wake-up control based on traffic prediction results. This makes it difficult to achieve high-efficiency network operation while ensuring the quality of communication services, thus hindering the achievement of green communication goals.

[0011] Therefore, there is an urgent need for a virtual cell multi-base station collaborative communication technology solution that can integrate network topology characteristics, have the ability to predict uncertainty, and achieve efficient and stable decision-making in large-scale multi-base station collaborative scheduling scenarios, so as to meet the operational requirements of future high-density, highly dynamic, and low-energy wireless communication networks. Summary of the Invention

[0012] The technical problem to be solved by the present invention is to overcome the shortcomings of existing wireless communication network control technology, such as lack of effective modeling of network space topology, inability to perceive prediction uncertainty leading to poor network robustness, and slow convergence of multi-base station cooperative scheduling algorithm which makes it difficult to meet real-time requirements. The present invention provides a flexible virtual cell multi-base station cooperative communication system and implementation method based on spatiotemporal prediction.

[0013] To solve the above-mentioned technical problems, the technical solution provided by the present invention is a flexible virtual cell multi-base station cooperative communication system based on spatiotemporal prediction, characterized in that:

[0014] The system comprises a five-layer architecture that is sequentially connected and forms a closed loop: a global data perception and graph construction layer, a spatiotemporal graph prediction layer, a flexible virtual cell networking layer, a prediction enhancement and collaborative communication layer, and a closed-loop feedback optimization layer. The global data perception and graph construction layer is used to collect network operation data and construct a dynamic interference topology map representing the interference relationship between base stations. The spatiotemporal graph prediction layer infers future traffic distribution based on the dynamic interference topology map and outputs a confidence index of the prediction results. The flexible virtual cell networking layer dynamically delineates the service boundaries of virtual cells based on the prediction results and the corresponding confidence index. The prediction enhancement and collaborative communication layer uses prediction information to perform collaborative resource scheduling and energy efficiency control for multiple base stations within the virtual cell. The closed-loop feedback optimization layer triggers closed-loop adaptive adjustment of the system by monitoring prediction deviations and network operation indicators.

[0015] As an improvement, the data collected by the global data perception and graph construction layer includes historical traffic sequences of base stations, mobile trajectories of user terminals, service slice demand identifiers, and channel state information. The dynamic interference topology graph uses base stations as nodes and path loss or interference coupling degree between base stations as the weight of the connection edges, transforming discrete network state data into graph structure data.

[0016] As an improvement, the spatiotemporal graph prediction layer adopts a spatiotemporal graph neural network model that includes a spatial graph convolution module and a temporal attention module. While outputting the predicted traffic value and hotspot drift vector for future time periods, the spatiotemporal graph neural network model simultaneously outputs the variance value, which represents the uncertainty of the prediction result, as a confidence index.

[0017] As an improvement, the elastic virtual cell networking layer determines the anchor base station of the virtual cell based on the hotspot drift vector and establishes a mapping relationship between the coverage area of ​​the virtual cell and the prediction confidence index. When the prediction confidence is higher than a preset threshold, the virtual cell boundary is automatically shrunk to achieve accurate coverage, and when the prediction confidence is lower than a preset threshold, the virtual cell boundary is automatically expanded to include redundant cooperative nodes.

[0018] As an improvement, the prediction-enhanced collaborative communication layer integrates a deep reinforcement learning scheduling model with an action masking mechanism. The action masking mechanism uses the active region information output by the spatiotemporal graph prediction layer to generate a mask matrix, which masks invalid base station or beam actions during the reinforcement learning scheduling process, thus limiting the search space for resource scheduling to hotspot areas.

[0019] As an improvement, the closed-loop feedback optimization layer includes a digital twin verification unit and a policy feedback unit. The digital twin verification unit is used to construct a simulation environment consistent with the existing network in the logical domain and to verify the candidate control strategies before the control strategy is issued to the physical network. The policy feedback unit is used to compare the deviation between the prediction results and the actual operating state. When the deviation exceeds a preset threshold, the scheduling parameters are corrected or the model is updated to achieve closed-loop adaptive optimization of the system.

[0020] A method for implementing multi-base station cooperative communication in elastic virtual cells based on spatiotemporal prediction includes the following steps:

[0021] Step 1: Collect multi-source heterogeneous network operation data through the open wireless access network interface, and perform preprocessing such as cleaning, alignment and standardization on the collected data. Construct a dynamic interference topology map based on the geographical location of the base station and the signal interference relationship.

[0022] Step 2: Input the dynamic interference topology map into the pre-trained spatiotemporal graph neural network model to deduce the traffic distribution status in a specific future time period, output the drift trajectory of the hotspot center, and simultaneously calculate the confidence variance of the prediction results;

[0023] Step 3: Select the anchor base station of the virtual cell based on the hotspot center drift trajectory, and dynamically adjust the coverage of the virtual cell according to the confidence variance to define the service boundary of the elastic virtual cell.

[0024] Step 4: Within the elastic virtual cell, an action mask is generated using the prediction results to limit the scheduling space, and collaborative resource scheduling and beamforming are performed on multiple base stations;

[0025] Step 5: Monitor key performance indicators and prediction deviations of the network in real time. When the deviation exceeds a preset threshold, trigger model parameter updates or schedule policy resets to form a closed-loop operation.

[0026] As an improvement, the dynamic interference topology map in step 1 uses base stations as nodes and the path loss or interference coupling degree between base stations as the weight of the connection edge. The interference coupling degree is calculated based on historical measurement data or real-time measurement reports and is used to characterize the actual interference intensity relationship between base stations.

[0027] As an improvement, the spatiotemporal graph neural network model in step 2 includes a spatial feature extraction module and a temporal feature modeling module. The spatial feature extraction module is used to characterize the spatial dependency relationship in the interference topology between base stations, and the temporal feature modeling module is used to extract the periodic and burst features of network traffic changing over time.

[0028] As an improvement, the prediction confidence variance in step 2 is obtained by performing multiple forward propagations on the same input during the inference phase. The confidence variance is used to characterize the level of uncertainty of the prediction result in the time and space dimensions.

[0029] As an improvement, the adjustment of the virtual cell coverage area in step 3 is based on the prediction confidence variance and the preset mapping relationship. When the prediction confidence variance is small, the virtual cell coverage area is reduced to reduce the cooperation overhead. When the prediction confidence variance is large, the virtual cell coverage area is expanded to introduce redundant cooperative base stations.

[0030] As an improvement, the step 4 of generating an action mask using the prediction results specifically includes: identifying inactive areas based on the predicted hotspot distribution information, generating a corresponding mask matrix, and using this mask matrix to shield base station activation or beam pointing actions targeting inactive areas when the reinforcement learning algorithm selects actions, thereby reducing the search space of the scheduling algorithm and improving decision-making efficiency.

[0031] As an improvement, the closed-loop operation in step 5 includes a digital twin verification process. Before the scheduling strategy or control parameters are sent to the physical network, the strategy is first simulated and verified in the digital twin environment of the logical domain. The strategy is only sent out when the verification results meet the preset performance and stability conditions.

[0032] The advantages of this invention compared to existing technologies are as follows: By constructing a dynamic interference topology map and introducing a spatiotemporal graph neural network for joint modeling, this invention overcomes the technical limitations of treating base stations as independent objects for traffic prediction. It can explicitly characterize the interference coupling relationship between base stations and the spatial correlation caused by user movement, thereby significantly improving the accuracy of hotspot drift prediction in complex interference environments and highly dynamic service scenarios. Furthermore, this invention introduces a prediction confidence index as an important decision-making basis for virtual cell networking. It dynamically adjusts the coverage range of virtual cells according to changes in prediction uncertainty and automatically introduces redundant cooperative base stations when prediction reliability is insufficient. This effectively reduces the risk of handover failures and insufficient coverage caused by rigid networking strategies in sudden scenarios, enhancing the robustness and fault tolerance of the system. Simultaneously, this invention uses prediction results to generate action masks to constrain the action space of reinforcement learning algorithms, effectively reducing the computational complexity of multi-base station collaborative scheduling problems. This enables scheduling strategies to converge quickly in large-scale scenarios, thereby meeting the millisecond-level real-time control requirements of next-generation wireless communication networks. Furthermore, this invention combines prediction results and their confidence information to implement tiered sleep control for low-load base stations. This achieves refined energy management for base stations in non-hotspot areas while ensuring service reliability, reducing unnecessary energy consumption and improving the overall network energy efficiency. Overall, the system architecture adopted in this invention is compatible with the concepts of open radio access networks and digital twin technology, possessing good engineering feasibility and compatibility. It can support the smooth deployment of existing networks and meet the needs of future wireless communication network evolution. Attached Figure Description

[0033] Figure 1 This is a schematic diagram of the multi-base station cooperative communication system architecture of the elastic virtual cell based on spatiotemporal prediction of the present invention.

[0034] Figure 2 This is a schematic diagram of the process for implementing multi-base station cooperative communication in an elastic virtual cell based on spatiotemporal prediction according to the present invention. Detailed Implementation

[0035] To facilitate understanding of this application, a more complete description will be provided below with reference to the accompanying drawings, which illustrate embodiments of the present application. However, the present application can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided so that the disclosure of this application will be thorough and complete.

[0036] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting of the application.

[0037] It is understood that spatial relation terms such as "below," "under," "below," "below," "above," "over," etc., can be used here to describe the relationship between one element or feature shown in the figure and other elements or features. It should be understood that, in addition to the orientation shown in the figure, spatial relation terms also include different orientations of the device in use and operation. For example, if the device in the figure is flipped, the element or feature described as "below" or "under" or "below" of the other element or feature will be oriented "over" the other element or feature. Therefore, the exemplary terms "below" and "under" can include both upper and lower orientations. Furthermore, the device may also include other orientations, such as being rotated 90 degrees or other orientations, and the spatial descriptive terms used herein will be interpreted accordingly.

[0038] It should be noted that when one element is considered to be "connected" to another element, it can be directly connected to the other element or connected to the other element through an intermediary element. In the following embodiments, "connection" should be understood as "electrical connection," "communication connection," etc., if the connected circuits, modules, units, etc., have the transmission of electrical signals or data between them.

[0039] When used herein, the singular forms of “a,” “an,” and “the” may also include the plural forms unless the context clearly indicates otherwise. It should also be understood that the terms “comprising,” “including,” or “having,” etc., specify the presence of the stated feature, whole, step, operation, component, part, or combination thereof, but do not preclude the possibility of the presence or addition of one or more other features, wholes, steps, operations, components, parts, or combinations thereof.

[0040] Combined with appendix Figure 1 A spatiotemporal prediction-based elastic virtual cell multi-base station cooperative communication system is proposed. This system comprises a multi-layered functional architecture with physical connections and logical closed-loop operation, used to achieve predictive cooperative communication control of multiple base stations. The system includes a global data perception and map construction layer, a spatiotemporal map prediction layer, an elastic virtual cell networking layer, a prediction-enhanced cooperative communication layer, and a closed-loop feedback optimization layer. These layers operate in a closed loop through the interaction of data and control flows. The global data perception and map construction layer acquires basic network operation status information; the spatiotemporal map prediction layer performs spatiotemporal joint analysis of dynamic interference topology maps; the elastic virtual cell networking layer dynamically constructs virtual cells based on prediction results; the prediction-enhanced cooperative communication layer performs radio resource scheduling and cooperative communication control within the virtual cells; and the closed-loop feedback optimization layer continuously monitors and optimizes the system's operational performance.

[0041] The input parameters of the global data perception and map construction layer include historical traffic data of base stations, mobile trajectory information of user terminals, service slice demand identifiers, and inter-cell interference measurement information. This layer preprocesses and structures the input data, and its output parameter is a dynamic interference topology map that characterizes the network topology. This topology map uses base stations as nodes and describes the spatial structural characteristics of the network by reflecting the connection relationships between base stations, which reflect the degree of interference or correlation between base stations.

[0042] The input parameters of the spatiotemporal graph prediction layer include the dynamic interference topology map and the corresponding time-series traffic characteristics. This layer employs a spatiotemporal graph neural network to comprehensively analyze network spatial relationships and traffic temporal evolution patterns. While outputting the traffic prediction results for each base station in the future time period, it simultaneously outputs the variance of the prediction results. This variance value is used as a measure of the uncertainty of the prediction results. Furthermore, this layer also outputs a hotspot drift vector, indicating the direction of movement of the hotspot center within the network topology.

[0043] The input parameters of the elastic virtual cell networking layer include traffic prediction results, prediction uncertainty indicators, and hotspot drift vectors. This layer determines the anchor base station of the virtual cell based on the predicted location of the hotspot area and dynamically adjusts the coverage area of ​​the virtual cell according to the prediction uncertainty indicator. This reduces the cooperation scale when prediction reliability is high and introduces redundant cooperative base stations to improve system stability when prediction uncertainty exists. Its output parameters are the set of base stations for the virtual cell and the corresponding coverage boundary parameters.

[0044] The input parameters of the prediction-enhanced collaborative communication layer include the set of base stations within the virtual cell, predicted hotspot area information, and real-time network status information. To improve scheduling efficiency, this layer employs a deep reinforcement learning scheduling model with an integrated action masking mechanism and uses the hotspot area information output by the prediction layer to limit the scope of scheduling decisions. During scheduling, joint scheduling is performed only on base stations, beams, and resource units related to hotspot areas. This layer outputs the resource allocation scheme, transmit power control parameters, and beam configuration parameters corresponding to each base station, thereby reducing the computational complexity of scheduling while ensuring communication quality.

[0045] The input parameters of the closed-loop feedback optimization layer include key performance indicators collected during actual network operation and predicted expected performance indicators, such as throughput, packet loss rate, and energy consumption. This layer analyzes the deviation between actual performance and predicted results, and uses digital twin technology to perform policy pre-validation in the logical domain. When the deviation exceeds a preset threshold, it outputs a prediction model update command or a virtual cell networking and scheduling policy adjustment command to drive parameter updates in each functional layer of the system, forming a closed-loop optimization process that combines prediction, control, and feedback.

[0046] Combined with appendix Figure 2 The present invention also provides a method for implementing multi-base station cooperative communication in elastic virtual cells based on spatiotemporal prediction, comprising the following steps:

[0047] Step 1: Collect multi-source heterogeneous data through the open radio access network interface. The collected data includes base station traffic statistics, user terminal mobility information, service demand identifiers, and inter-base station interference measurement information. The system cleans, normalizes, and structures the data, and constructs a dynamic interference topology map based on the interference relationships between base stations to characterize the network spatial structure.

[0048] Step 2: Input the dynamic interference topology map and the corresponding historical traffic time series data into the pre-trained spatiotemporal graph neural network model. The prediction results include the traffic distribution status in the future time period, the location change trend information of hotspot areas, and the corresponding prediction uncertainty index.

[0049] Step 3: Determine the anchor base station for the virtual cell based on the location change trend information of the hotspot area, and dynamically determine the coverage area of ​​the virtual cell by combining the prediction uncertainty index. This outputs the current virtual cell networking result, including the set of participating base stations and their coverage boundary parameters.

[0050] Step 4: Using the virtual cell networking results, predicted hotspot activity area information, and real-time network status information as input, a deep reinforcement learning algorithm integrating an action masking mechanism is used to constrain and optimize the wireless resource scheduling process. The scheduling results include the joint configuration scheme of each base station in terms of time-frequency resources, transmit power, and beam direction.

[0051] Step 5: Monitor network performance data in real time and compare actual performance indicators with predicted results. When a deviation exceeds a preset threshold, the system triggers an update to the prediction model parameters or an adjustment to the virtual cell networking and scheduling strategy, thereby achieving online optimization of the prediction model and control strategy and forming a closed-loop adaptive control process.

[0052] The spatiotemporal graph neural network model described in step 2 enables early prediction of the spread trend of burst traffic in the network topology by jointly modeling the spatial dependencies and temporal evolution patterns between base station nodes.

[0053] The joint scheduling in step 4 aims to maximize the system weighted sum rate and minimize the overall network energy consumption. For areas with low prediction confidence, the scheduling strategy automatically increases the allocation ratio of control channel resources to ensure connection reliability.

[0054] Step 4 also includes a prediction-guided sleep control mechanism: for base stations with low predicted traffic and high confidence, a deep sleep command is issued in advance; for base stations with low predicted traffic but insufficient confidence, a shallow sleep or standby state is maintained to cope with potential sudden service demands.

[0055] The closed-loop feedback optimization layer described in step 5 uses digital twin technology to perform parallel simulation in the logic domain to verify the stability of the optimization strategy before sending it down to the physical network for execution.

[0056] I. Implementation Background and Overall Approach:

[0057] In existing 5G and future 6G networks, user service demands are changing more frequently in time and space, resulting in network traffic exhibiting significant non-stationarity and burstiness. In scenarios such as urban commercial districts, transportation hubs, and sports stadiums, localized areas are prone to becoming communication hotspots during specific time periods, leading to increased load and decreased communication performance on related base stations, while adjacent base station resources remain underutilized. Traditional network control schemes based on fixed thresholds or simple time-series predictions struggle to accurately reflect the topological relationships between base stations and lack effective mechanisms to address prediction uncertainties, easily leading to communication quality degradation or handover failures in bursty traffic scenarios.

[0058] Based on the above background, this implementation constructs a dynamic interference topology map, incorporating the interference relationships and spatial correlations between base stations into a unified modeling process. It also introduces a spatiotemporal graph neural network to predict and analyze the evolution trend of network traffic, thereby improving the accuracy of hotspot location change prediction. On this basis, the system dynamically constructs elastic virtual cells using the prediction results and their uncertainties, allowing the coverage area of ​​the virtual cells to be adjusted according to the prediction reliability. Subsequently, the system collaboratively schedules resources from multiple base stations within the virtual cells and reduces the computational complexity of scheduling through a prediction-guided action masking mechanism, improving scheduling real-time performance. Simultaneously, the system continuously monitors the actual network operating status and adjusts the prediction model and scheduling strategy through a closed-loop feedback mechanism, thus forming an overall technical solution combining prediction, networking, scheduling, and feedback. This solution is suitable for high-density, highly dynamic user scenarios, helping to improve network resource utilization efficiency, reduce network energy consumption, and support the intelligent evolution of wireless communication networks.

[0059] II. Specific Implementation at Each Level of the System:

[0060] (I) Specific Implementation Methods of the Global Data Perception and Map Construction Layer

[0061] In this embodiment, the global data perception and graph construction layer is deployed based on an open wireless access network architecture. It is used to uniformly collect and structure multi-source heterogeneous operational data in the wireless access network and convert it into dynamic graph data that can be used for subsequent intelligent inference.

[0062] In its implementation, this layer achieves cross-protocol stack data awareness through standard interfaces defined in the Open Radio Access Network. Physical layer data is collected by the Open Radio Unit, including reference signal received power, signal-to-noise ratio (SNR), and channel state information, reflecting the wireless propagation environment and link quality. Scheduling and buffering data is collected by the Open Distributed Unit, including buffer queue length, automatic retransmission feedback statistics, and scheduling resource block utilization, characterizing the base station's load and scheduling pressure. Furthermore, user terminal movement trajectory information and service slice identifiers are collected from each base station node via the control interface, reflecting the user's spatial distribution and service demand characteristics within the network.

[0063] After completing the above data collection, the system maps the physical network as a graph structure that evolves over time:

[0064]

[0065] Among them, the node set Each node in the database corresponds to a base station. The feature vector of each node... The feature vector is used to describe the operating status of the base station at the current moment, and includes at least the base station load rate, the number of active users, and the average queuing delay.

[0066] edge set This is used to characterize the coupling relationship between base stations. The coupling relationship is no longer defined solely based on geographical distance, but is modeled according to the interference intensity between base stations. The system calculates the path loss matrix between base stations and the edge weights based on measurement reports or drive test data. It is inversely proportional to path loss, thus reflecting the degree of potential interference between base stations.

[0067] Adjacency Matrix The system is updated at a preset period to reflect the dynamic changes in network topology and interference relationships. The completed dynamic interference topology map is transmitted as a data stream to the intelligent control node of the non-real-time wireless access network, providing structured input for subsequent spatiotemporal simulations.

[0068] (II) Specific Implementation Methods of Spatiotemporal Map Deduction and Prediction Layer

[0069] In this embodiment, the spatiotemporal graph extrapolation and prediction layer is deployed in the intelligent control node of the non-real-time wireless access network to perform joint spatial and temporal modeling of the dynamic interference topology and extrapolate the evolution trend of network traffic.

[0070] This layer uses a spatiotemporal graph convolutional network to process the input topological graph sequence. The model structure consists of alternating spatial feature extraction modules and temporal feature extraction modules.

[0071] In the spatial dimension, graph convolution operations are used to model the spatial dependencies between base station nodes. The feature update process of a layer is represented as follows:

[0072]

[0073] in, This represents the adjacency matrix after adding self-connections. This is the corresponding degree matrix. These are learnable weight parameters. This is a non-linear activation function. Using the above method, the model can capture the propagation relationship of base station traffic changes in the interference topology.

[0074] In the time dimension, the model models the node feature sequence along the time axis to extract the periodic features and sudden change features of traffic evolution over time.

[0075] To improve the reliability of prediction results in network control, this embodiment introduces a Monte Carlo random deactivation inference mechanism in the model output stage. That is, the random deactivation layer is kept on and executed multiple times during the inference stage (e.g., ...). (Next) forward propagation, and then by calculating the statistical distribution of these predictions, The mean of the predictions As the most likely traffic load for the base station in the future, while also considering the variance As a confidence index for quantifying uncertainty (variance) The larger the value, the higher the uncertainty, thus enabling an accurate assessment of the reliability of the prediction results.

[0076] Based on the predicted flow heatmap sequence, the system further calculates the displacement of the geometric centroid of the hotspot region over continuous time slices, generating a hotspot drift vector. The vector reflects both the direction and speed of hotspot movement, and is used to guide the selection of virtual cell anchor points at the next level.

[0077] (III) Specific Implementation Methods of the Elastic Virtual Cell Networking Layer

[0078] In this embodiment, the elastic virtual cell networking layer is deployed in the near real-time wireless access network intelligent control node, and is used to dynamically construct and adjust the virtual cell structure based on the prediction results.

[0079] Based on the hotspot drift vector output by the spatiotemporal map prediction layer, the system predicts the location of the hotspot centroid at the next moment and selects the base station closest to the predicted location as the anchor base station for the virtual cell. Around the anchor base station, the system constructs the initial coverage area of ​​the virtual cell.

[0080] The coverage radius of a virtual cell is related to prediction uncertainty. The system establishes the coverage radius. With prediction variance The nonlinear mapping relationship between them is calculated as follows:

[0081]

[0082] in, This represents the minimum radius required to meet basic coverage needs. and This is the adjustment coefficient.

[0083] When the prediction variance is small and the prediction results are highly reliable, the system automatically shrinks the coverage area of ​​the virtual cell, including only the base stations related to the core hotspot in the cooperative set, in order to reduce coordination overhead and improve system energy efficiency. When the prediction variance is large and the prediction results are uncertain, the system automatically expands the coverage radius, including more peripheral neighboring nodes in the virtual cell as a "protective belt". Even if the hotspot drifts unexpectedly, the redundant base stations can immediately take over the service, thus trading resources for robustness.

[0084] The final set of cooperative base stations for the virtual cell is sent to the underlying scheduling entity through the control interface to complete the logical construction of the virtual cell.

[0085] (iv) Specific implementation of the prediction enhancement cooperative communication layer

[0086] In this embodiment, the prediction-enhanced cooperative communication layer is used to perform radio resource scheduling and cooperative communication control within the virtual cell. To address the issue of excessively high scheduling dimensionality in multi-base station and massive MIMO scenarios, this embodiment employs a deep reinforcement learning algorithm based on action masks and introduces a zero-trust sleep mechanism.

[0087] The system employs a near-end strategy optimization algorithm for joint resource scheduling. Its state space includes the channel quality of users within the virtual cell, the buffer state, and the predicted future traffic change trend.

[0088] To address the issue of extremely slow convergence in traditional reinforcement learning when searching for actions across all base stations and beams in the entire network, this embodiment introduces a crucial action masking mechanism. This mechanism utilizes active region information from the prediction layer output to generate a mask matrix. Specifically, if it is predicted that no user will exist in a certain beam direction in the future time slot, the corresponding mask will be set to 0.

[0089] After the policy network outputs the action probability distribution, the system first compares it with the mask matrix. Element-wise multiplication is performed, followed by normalization. This operation forces the agent to explore only within the subspace where the prediction is valid, thus significantly reducing the computational complexity of scheduling and enabling rapid decision-making. Based on this, the reward function of the scheduling algorithm is defined as:

[0090]

[0091] in Represents the total reward value; Represents the total user speed within the virtual cell; Represents the end-to-end latency of the service; This represents the energy consumption of the base station; , and These are the dynamic weighting coefficients for the three indicators mentioned above, used to balance the optimization needs under different business scenarios.

[0092] In terms of energy efficiency control, the system introduces a sleep control mechanism based on prediction results. Instead of passively waiting for base stations to be completely idle before shutting down, the system makes judgments based on prediction results. Specifically, the judgment logic is as follows: when the predicted traffic of a base station is lower than a preset threshold and the prediction variance is small, the system triggers deep sleep control; when the predicted traffic is low but the prediction uncertainty is high, the base station remains in a listening state to be woken up at any time, thus balancing energy-saving requirements with the ability to respond to sudden service disruptions.

[0093] (V) Specific Implementation Methods of the Closed-Loop Feedback Optimization Layer

[0094] In this embodiment, the closed-loop feedback optimization layer is used to construct the adaptive control closed loop of the system to ensure the stability and reliability of the system in the long term.

[0095] Specifically, during network operation, each base station continuously reports key performance indicators, including actual throughput, handover success rate, energy consumption level, and prediction deviation. The system analyzes these real-time performance indicators and adjusts scheduling parameters through a feedback mechanism when it detects performance degradation caused by scheduling policies.

[0096] Based on this, the system implements a parallel optimization strategy for both short-term and long-term scenarios. At the short-term closed-loop level, if real-time monitoring results indicate that the current scheduling strategy is causing performance degradation, such as a sudden increase in packet loss rate, the reinforcement learning agent will receive a negative feedback signal and quickly adjust strategy parameters such as beam power weights using an online learning algorithm. At longer timescales, the system continuously performs statistical analysis on prediction errors. Once it is detected that the spatiotemporal graph convolutional network model has consistently failed to predict accurately in a specific region, such as due to changes in pedestrian flow patterns caused by new road construction, the system immediately triggers a model retraining process. This process updates the neural network weights using the latest historical data and redeploys the optimized model to the inference module.

[0097] It is worth emphasizing that, to ensure the security of the existing network operation, the system first simulates and verifies the new policy in a digital twin environment in the logical domain before officially distributing the control policy to the physical base stations. Only when the new policy scores better than the current policy in the simulation environment and meets the preset requirements in terms of stability and performance indicators is it allowed to be distributed to the physical network for execution via the interface.

[0098] III. Detailed Explanation of Implementation Methods and Steps:

[0099] Step 1: Global Data Perception and Dynamic Map Construction

[0100] In this embodiment, firstly, based on the standard interface defined in the open radio access network architecture, multi-source heterogeneous operational data of the radio access network are uniformly collected. The collected data includes physical layer channel state information, scheduling and buffering related statistics, and user terminal movement trajectories and service slice identifiers. The system cleans and spatiotemporally aligns the above data to eliminate differences in sampling periods and formats between different data sources, and maps base stations in the network as nodes in a graph structure, and maps the interference coupling relationships between base stations as connection edges and their weights, thereby constructing a dynamic interference topology graph that reflects the real-time spatial structure characteristics of the network.

[0101] Step 2: Spatiotemporal mapping and prediction uncertainty assessment

[0102] After constructing the dynamic network topology, the sequence of dynamic interference topology maps is input into a pre-trained spatiotemporal graph convolutional network model to jointly model the spatial diffusion characteristics of network traffic within the topology and its evolutionary trend over time. While projecting traffic distribution over future periods, the model incorporates a random deactivation inference mechanism to perform multiple forward computations, obtaining the mean and variance of the prediction results through statistical analysis. The prediction mean characterizes the evolution trend of hotspot regions, while the prediction variance quantifies the uncertainty level of the prediction results, thereby generating a hotspot drift vector to describe the direction and trend of hotspot movement.

[0103] Step 3: Dynamically constructing flexible virtual cells

[0104] Based on the hotspot drift vector output in step 2, the system predicts the spatial location of the hotspot area at the next moment and selects the base station closest to the predicted location as the anchor base station for the virtual cell. Furthermore, the system dynamically adjusts the coverage area of ​​the virtual cell based on the prediction uncertainty. When the prediction uncertainty is low, the virtual cell range is automatically shrunk, including only base stations highly correlated with the hotspot to reduce coordination overhead and improve energy efficiency. When the prediction uncertainty is high, the coverage area is automatically expanded, introducing more neighboring base stations to form a redundant cooperative set, thereby enhancing the system's adaptability to sudden traffic changes. The final determined base station cooperative set is distributed through the control interface, completing the logical construction of the virtual cell.

[0105] Step 4: Predictive Enhancement of Cooperative Communication Scheduling and Energy Efficiency Control

[0106] After the virtual cell is constructed, the system performs multi-base station collaborative communication scheduling within the virtual cell. Edge-side control nodes employ a near-end policy optimization algorithm to jointly optimize radio resources and generate an action constraint matrix using hotspot distribution information obtained during the prediction phase. Base stations, beam directions, or resource units without service demand are masked, thereby limiting the scheduling search space and making decisions only within the predicted active areas. Simultaneously, the system implements a confidence-based sleep control strategy based on the prediction results. Base stations with low predicted traffic and low uncertainty are subject to deep sleep control, while base stations with low predicted traffic but high uncertainty are maintained in a shallow sleep state, achieving a balance between energy efficiency optimization and service responsiveness.

[0107] Step 5: Digital Twin Verification and Closed-Loop Feedback Optimization

[0108] During operation, the system continuously collects key network performance indicators and compares and analyzes the actual results with the predicted results. Before the scheduling or networking strategy is officially issued, the system first simulates and verifies the candidate strategies in a digital twin environment. Only when the verification results meet the stability and performance requirements are the relevant strategies allowed to be executed. When a decline in network performance or a long-term deviation of the predicted results from the actual situation is detected, the system triggers a closed-loop optimization mechanism based on the degree of deviation. This mechanism achieves adaptive evolution in response to changes in the network environment by adjusting scheduling parameters in the short term or updating the prediction model in the long term.

[0109] Beneficial effects:

[0110] The AI-based predictive virtual cell multi-base station collaborative wireless access network control system achieves deep integration of network perception, prediction, and control. It breaks through the traditional wireless access network's reliance on static configuration and passive response, enabling the network to predict and coordinate resource allocation before changes in service demands. This significantly improves the network's timeliness in responding to hotspot changes, the accuracy of resource allocation, and the overall stability of system operation.

[0111] The global data perception and dynamic graph modeling mechanism integrates and structures operational data scattered across different protocol layers and network nodes, transforming network status from traditional discrete performance index descriptions into a graph form that reflects spatial relationships and interference structures. This allows for a more accurate characterization of the real-time operational characteristics of wireless access networks, providing a stable and reliable data foundation for subsequent prediction and control.

[0112] The spatiotemporal mapping model jointly models the spatial diffusion characteristics of network traffic in the topology and its evolution over time. This enables hotspot prediction to comprehensively consider the mutual influence between base stations, rather than relying solely on a single base station or single time series analysis. This improves the accuracy of hotspot prediction and enhances the ability to perceive hotspot migration trends in advance.

[0113] The prediction uncertainty quantification mechanism differentiates and evaluates prediction reliability while outputting prediction results, enabling the system to identify prediction scenarios under different confidence levels. This avoids relying directly on a single prediction result for network control decisions under high uncertainty conditions, thereby effectively reducing the risk of misjudgment and improving the stability and robustness of network operation.

[0114] The uncertainty-aware virtual cell networking scheme based on prediction results enables dynamic adjustment of the virtual cell coverage. When the prediction is reliable, it accurately focuses on hotspot areas, reduces the participation of unnecessary cooperative base stations, lowers cooperative communication overhead, and improves system energy efficiency. When the prediction is uncertain, it introduces redundant cooperative nodes to form a protection zone, thereby enhancing the system's capacity to withstand sudden traffic changes and its fault tolerance.

[0115] The predictive-enhanced collaborative scheduling strategy effectively constrains the scheduling action space by utilizing hotspot distribution information, focusing scheduling decisions on potentially active areas, significantly reducing scheduling search complexity and shortening decision time, thereby meeting the requirements of wireless access networks for real-time scheduling response capabilities.

[0116] The energy efficiency control system based on prediction confidence distinguishes and manages base station operation modes under different load states and different prediction reliability conditions. It achieves dynamic optimization of base station energy consumption while ensuring service continuity, thereby achieving a coordinated balance between energy efficiency improvement and service quality assurance at the overall network level.

[0117] The digital twin-driven closed-loop feedback optimization architecture performs simulation verification of candidate strategies before the control strategy is issued to the live network, reducing the potential risks of strategy adjustment to the actual network operation. By continuously comparing the prediction results with the actual operating status, it triggers the correction of scheduling parameters or model updates, enabling the system to adapt to long-term changes in the network environment and improving the overall operational reliability and continuous optimization capability.

[0118] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present invention; the actual structure is not limited thereto. In conclusion, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the invention, such designs should fall within the protection scope of the present invention.

Claims

1. A flexible virtual cell multi-base station cooperative communication system based on spatiotemporal prediction, characterized in that: The system comprises a five-layer architecture that is sequentially connected and forms a closed loop: a global data perception and graph construction layer, a spatiotemporal graph prediction layer, a flexible virtual cell networking layer, a prediction enhancement and collaborative communication layer, and a closed-loop feedback optimization layer. The global data perception and graph construction layer is used to collect network operation data and construct a dynamic interference topology map representing the interference relationship between base stations. The spatiotemporal graph prediction layer infers future traffic distribution based on the dynamic interference topology map and outputs a confidence index of the prediction results. The flexible virtual cell networking layer dynamically delineates the service boundaries of virtual cells based on the prediction results and the corresponding confidence index. The prediction enhancement and collaborative communication layer uses prediction information to perform collaborative resource scheduling and energy efficiency control for multiple base stations within the virtual cell. The closed-loop feedback optimization layer triggers closed-loop adaptive adjustment of the system by monitoring prediction deviations and network operation indicators.

2. The elastic virtual cell multi-base station cooperative communication system based on spatiotemporal prediction according to claim 1, characterized in that: The data collected by the global data perception and graph construction layer includes historical traffic sequences of base stations, mobile trajectories of user terminals, service slice demand identifiers, and channel state information. The dynamic interference topology graph uses base stations as nodes and path loss or interference coupling between base stations as the weight of the connection edges, transforming discrete network state data into graph structure data.

3. The elastic virtual cell multi-base station cooperative communication system based on spatiotemporal prediction according to claim 1, characterized in that: The spatiotemporal graph prediction layer adopts a spatiotemporal graph neural network model that includes a spatial graph convolution module and a temporal attention module. While outputting the predicted traffic value and hotspot drift vector for future time periods, the spatiotemporal graph neural network model simultaneously outputs the variance value, which represents the uncertainty of the prediction result, as a confidence index.

4. The elastic virtual cell multi-base station cooperative communication system based on spatiotemporal prediction according to claim 1, characterized in that: The elastic virtual cell networking layer determines the anchor base station of the virtual cell based on the hotspot drift vector and establishes a mapping relationship between the coverage area of ​​the virtual cell and the prediction confidence index. When the prediction confidence is higher than a preset threshold, the virtual cell boundary is automatically shrunk to achieve accurate coverage. When the prediction confidence is lower than a preset threshold, the virtual cell boundary is automatically expanded to include redundant cooperative nodes.

5. The elastic virtual cell multi-base station cooperative communication system based on spatiotemporal prediction according to claim 4, characterized in that: The prediction-enhanced collaborative communication layer integrates a deep reinforcement learning scheduling model with an action masking mechanism. The action masking mechanism uses the active region information output by the spatiotemporal graph prediction layer to generate a mask matrix, which masks invalid base station or beam actions during the reinforcement learning scheduling process, thus limiting the search space for resource scheduling to hotspot areas.

6. The elastic virtual cell multi-base station cooperative communication system based on spatiotemporal prediction according to claim 1, characterized in that: The closed-loop feedback optimization layer includes a digital twin verification unit and a policy feedback unit. The digital twin verification unit is used to construct a simulation environment consistent with the existing network in the logical domain and to verify the candidate control strategies before the control strategy is issued to the physical network. The policy feedback unit is used to compare the deviation between the prediction results and the actual operating state. When the deviation exceeds a preset threshold, the scheduling parameters are corrected or the model is updated to achieve closed-loop adaptive optimization of the system.

7. A method for implementing multi-base station cooperative communication in elastic virtual cells based on spatiotemporal prediction, characterized in that: Includes the following steps: Step 1: Collect multi-source heterogeneous network operation data through the open wireless access network interface, and perform preprocessing such as cleaning, alignment and standardization on the collected data. Construct a dynamic interference topology map based on the geographical location of the base station and the signal interference relationship. Step 2: Input the dynamic interference topology map into the pre-trained spatiotemporal graph neural network model to deduce the traffic distribution status in a specific future time period, output the drift trajectory of the hotspot center, and simultaneously calculate the confidence variance of the prediction results; Step 3: Select the anchor base station of the virtual cell based on the hotspot center drift trajectory, and dynamically adjust the coverage of the virtual cell according to the confidence variance to define the service boundary of the elastic virtual cell. Step 4: Within the elastic virtual cell, an action mask is generated using the prediction results to limit the scheduling space, and collaborative resource scheduling and beamforming are performed on multiple base stations; Step 5: Monitor key performance indicators and prediction deviations of network operation in real time. When the deviation exceeds the preset threshold, trigger model parameter updates or scheduling strategy resets to form a closed-loop operation.

8. The method for implementing multi-base station cooperative communication in elastic virtual cells based on spatiotemporal prediction according to claim 7, characterized in that: The dynamic interference topology map described in step 1 uses base stations as nodes and the path loss or interference coupling degree between base stations as the weight of the connection edges. The interference coupling degree is calculated based on historical measurement data or real-time measurement reports and is used to characterize the actual interference intensity relationship between base stations.

9. The method for implementing multi-base station cooperative communication in elastic virtual cells based on spatiotemporal prediction according to claim 7, characterized in that: The spatiotemporal graph neural network model described in step 2 includes a spatial feature extraction module and a temporal feature modeling module. The spatial feature extraction module is used to characterize the spatial dependencies in the interference topology between base stations, and the temporal feature modeling module is used to extract the periodic and burst features of network traffic over time. The prediction confidence variance mentioned in step 2 is obtained by performing multiple forward propagations on the same input during the inference phase. The confidence variance is used to characterize the level of uncertainty of the prediction results in the time and space dimensions.

10. The method for implementing multi-base station cooperative communication in elastic virtual cells based on spatiotemporal prediction according to claim 7, characterized in that: The adjustment of the virtual cell coverage area in step 3 is based on the prediction confidence variance and the preset mapping relationship. When the prediction confidence variance is small, the virtual cell coverage area is reduced to reduce the cooperation overhead. When the prediction confidence variance is large, the virtual cell coverage area is expanded to introduce redundant cooperative base stations. Step 4, which involves generating an action mask using the prediction results, specifically includes: identifying inactive areas based on the predicted hotspot distribution information, generating a corresponding mask matrix, and using this mask matrix to block base station activation or beam pointing actions targeting inactive areas when the reinforcement learning algorithm selects actions, thereby reducing the search space of the scheduling algorithm and improving decision-making efficiency.

Images