Inter-cell interference coordination and load balancing method for dense heterogeneous network based on graph neural network

Abstract

The invention relates to the technical field of communication systems, in particular to a dense heterogeneous network inter-cell interference coordination and load balancing method based on a graph neural network, which comprises the following steps of: taking a user terminal as a first type of node; taking the base station as a second type node; constructing a heterogeneous edge based on the user-base station association relationship and the interference relationship between the base stations; outputting a heterogeneous network diagram fusing the spatial-temporal characteristics of the service; inputting the heterogeneous network graph into a dual-channel graph neural network model, wherein the dual-channel graph neural network model comprises an interference coordination channel and a load balancing channel; performing cross-base-station orthogonal distribution on the frequency spectrum units with the gain values higher than a preset gain threshold value; and migrating the users to the first available base station in the target base station sequence in batches according to the priority. According to the method, spectrum coupling weight is introduced through spectrum sensing graph convolution to calculate frequency point level interference reduction potential, user graph embedding, bandwidth fluctuation ratio and spatial information are fused based on an attention mechanism to generate a migration path, and bidirectional mining of multi-source heterogeneous features is realized.

Description

Technical Field

[0001] The present invention relates to the technical field of communication systems, and in particular to a method for interference coordination and load balancing between cells in dense heterogeneous networks based on graph neural networks. Background Art

[0002] With the continued evolution of 5G and subsequent generations of mobile communication networks, dense heterogeneous networks are being widely deployed as a core architecture to improve system capacity, coverage, and service quality. In this architecture, different types of base stations (such as macro base stations, micro base stations, and pico base stations) are deployed in an irregular and dense manner in urban environments to support diverse service scenarios (such as eMBB, URLLC, and mMTC). However, this highly dense and heterogeneous deployment structure also brings significant interference and load management challenges: On the one hand, due to the frequent coverage overlap of multi-level base stations in physical space, strong interference is easily caused during the spectrum resource reuse process, especially in hot spots with high spectrum reuse rates. Frequency conflicts will seriously weaken the spectrum efficiency of the system. Existing interference coordination methods mostly rely on static spectrum division or neighboring area interference maps, but it is difficult to perceive the dynamic impact of actual service distribution on the interference structure, and lack the ability to fine-grainedly identify the root causes of spectrum conflicts.

[0003] Furthermore, service traffic exhibits highly uneven characteristics across user space and time, leading to unbalanced load distribution across base stations. Traditional solutions often rely on single-factor heuristics such as signal strength and average load, failing to comprehensively consider user service characteristics, target base station resource constraints, and migration link costs. This results in inaccurate migration decisions, which can easily lead to migration failures or load rebound, impacting overall network stability.

[0004] In recent years, graph neural networks (GNNs), as a structured relationship modeling tool, have gained increasing attention in communication network optimization tasks. However, most existing graph learning-based resource scheduling methods model the network as a homogeneous graph structure, which cannot simultaneously capture the service coupling relationship between users and base stations and the interference relationship between base stations. This limits the model's ability to represent dynamic information from multiple sources. Summary of the Invention

[0005] The present invention provides a method for interference coordination and load balancing between cells in dense heterogeneous networks based on graph neural networks, a heterogeneous graph modeling method that integrates business, spatial and spectrum information, and combines it with a dual-channel graph neural network structure to implement a systematic solution for collaborative optimization of spectrum interference coordination and load balancing, thereby improving the resource utilization efficiency and service quality assurance capabilities of dense heterogeneous networks.

[0006] The method for interference coordination and load balancing between cells in dense heterogeneous networks based on graph neural networks includes the following steps: S1. Constructing a heterogeneous network graph with enhanced business features: The user terminal is regarded as the first type of node; the base station is regarded as the second type of node; based on the user-base station association relationship and the interference relationship between base stations, heterogeneous edges are constructed, wherein the heterogeneous edges include user-base station association edges and inter-base station interference edges; Output a heterogeneous network graph that integrates the spatiotemporal characteristics of services; S2. Dual-channel graph neural network joint decision-making: The heterogeneous network graph is input into a dual-channel graph neural network model, including an interference coordination channel and a load balancing channel, wherein: Interference coordination channel: Through the spectrum perception graph convolution layer, it generates the spectrum efficiency gain distribution matrix of each base station, indicating the theoretical interference reduction range of each spectrum unit; Load balancing channel: Generates a load pressure transmission path through the service migration graph attention layer, identifies the migration priority of the users to be migrated and the target base station sequence; S3. Perform spectrum-load joint reconstruction: performing cross-base station orthogonal allocation on spectrum units having gain values ​​higher than a preset gain threshold according to the spectrum efficiency gain distribution matrix; Along the load pressure transmission path, users are migrated in batches according to priority to the first available base station in the target base station sequence.

[0007] Optionally, the node attributes of the first type of node include service type, location coordinates, and a historical traffic fluctuation coefficient, where the historical traffic fluctuation coefficient is calculated based on a standard deviation of traffic over multiple time slots in the past.

[0008] Optionally, the node attributes of the second type of nodes include base station height, remaining fronthaul bandwidth, and available spectrum units.

[0009] Optionally, the edge weight of the user-base station association edge is calculated based on a weighted coupling of the user service type and the link quality indicator. By using a preset service type priority mapping function, the QoS levels corresponding to different service types are converted into numerical weights, and combined with the real-time signal-to-noise ratio between the user and the home base station to obtain a link quality indicator oriented to service performance; The edge weight of the inter-base station interference edge is calculated based on the product of spatial overlap and spectrum overlap. The overlap of the coverage areas is quantified by calculating the ratio of the number of shared service users between the two base stations to the total number of covered users. The degree of sharing of spectrum resources is calculated by the intersection-over-combination ratio of the available spectrum sets of the two base stations. The edge weight of the inter-base station interference edge reflects the interference intensity caused by the coupling of spatial topology and spectrum resources. The final output heterogeneous network graph includes the user node set, the base station node set, the user-to-base station association edge set, the interference edge set between base stations, the user node attribute matrix, the base station node attribute matrix, and the edge weight matrix.

[0010] Optionally, the interference coordination channel is used to model the spectrum interference relationship between base stations, and the spectrum perception graph convolution mechanism is used to perform dual-source neighborhood feature aggregation on the base station node, and the information from the adjacent user nodes and the adjacent base station nodes is integrated. The user node aggregation weight is generated based on the attention function of the service perception, and the base station node aggregation weight is calculated based on the spectrum overlap degree and the interference edge weight. After feature aggregation, the spectrum efficiency gain prediction value of each base station on each available spectrum unit is generated. , indicating a base station In spectrum unit The theoretical interference reduction amplitude on the spectrum efficiency gain is obtained, and the spectrum efficiency gain prediction value is combined with the base station hidden state, spectrum ownership identifier and the strongest interference link information, and is output through linear prediction and activation function; all spectrum efficiency gain prediction values ​​constitute a spectrum efficiency gain distribution matrix.

[0011] Optionally, the load balancing channel models the feasibility and priority of user migration, and based on the graph attention mechanism, constructs a migration utility scoring function between each user and the candidate target base station to obtain a migration utility score. The migration utility score integrates the graph embedding state of the user and the base station, the remaining forward transmission bandwidth of the target base station, the user traffic fluctuation coefficient and the spatial distance; based on the load rate and service type priority of the base station to which the user is currently connected, the migration priority is calculated, and the set of users to be migrated is screened.

[0012] Optionally, the load balancing channel also includes generating an ordered target base station sequence for each user to be migrated based on the migration utility score and hop attenuation mechanism, constructing a load pressure conduction path, and outputting a path set including user identification, migration priority and target base station sequence for load balancing scheduling.

[0013] Optionally, S3 includes spectrum unit screening, specifically including calculating the global average value of the spectrum efficiency gain matrix, and combining the adjustable coefficient to obtain the preset gain threshold under the current time slot, screening the spectrum efficiency gain value of each base station, and only retaining the spectrum units greater than the preset gain threshold to form a high-value spectrum unit set of the current base station as a candidate resource for priority scheduling.

[0014] Optionally, the cross-base station orthogonal allocation includes taking all base stations as nodes. If two base stations have overlapping frequency points in the high-gain spectrum set, a connection edge is established between them to form a spectrum conflict graph, and a graph coloring operation is performed on the conflict graph. All base stations are grouped, and it is required that there is no spectrum conflict between the base stations in each group. The respective high-gain spectrum units in each group are uniformly classified into the same orthogonal resource pool to ensure that adjacent base stations do not transmit at the same frequency point, thereby achieving interference avoidance and efficient reuse of spectrum resources.

[0015] Optionally, the S3 further includes: Migration batch division: All users to be migrated are sorted from high to low according to migration priority, with priority given to users in high-service-level or high-load areas. The number of users that can be accommodated in each migration batch is calculated based on the available fronthaul bandwidth of the current network and the average bandwidth requirements of a single user. All sorted users are then divided into several batches, and migration is carried out gradually in batches. Target base station selection: For each user in each batch, the pre-generated target base station sequence is traversed, and the resource status of each candidate base station is checked in turn. If the remaining fronthaul bandwidth of a base station exceeds the set safety threshold of the user's bandwidth demand and the number of its available spectrum units is at least twice the user's demand, the current base station is considered to have sufficient carrying capacity and is immediately marked as the user's migration target. Migration operation: allocates corresponding spectrum resources to the user in the orthogonal spectrum pool of the target base station, and completes the transfer of user context and data plane switching through the interface.

[0016] Beneficial effects of the present invention: The present invention constructs a heterogeneous network graph that integrates user type, location, traffic fluctuation, base station bandwidth, and spectrum status, and introduces user-base station service weights and inter-base station interference edge weights to achieve structured modeling of the relationship between service demand, channel status, and spectrum resources in the network. Compared with the traditional graph construction method based only on geographical distance or RSRP, the proposed graph structure significantly improves the graph neural network's ability to express interference propagation paths and load distribution patterns, providing a precise and dynamic network perception foundation for subsequent joint optimization.

[0017] This paper designs interference coordination and load balancing channels within a graph neural network architecture. The former uses spectrum-aware graph convolution to introduce spectrum coupling weights to calculate frequency-level interference reduction potential, while the latter uses an attention mechanism to fuse user graph embeddings, bandwidth fluctuation ratios, and spatial information to generate migration paths, enabling bidirectional mining of multi-source heterogeneous features. While the two channels model different tasks, they share the underlying graph structure and embedding space, ultimately forming an explicit linkage between spectrum allocation and user migration strategies, overcoming the limitations of traditional approaches that separate modeling of spectrum optimization and load migration.

[0018] During the execution phase, the present invention uses a spectrum screening and graph coloring orthogonal allocation method driven by gain thresholds to accurately identify and isolate conflicting frequencies, thereby reducing co-channel interference between cells. At the same time, during the user migration process, a target base station selection strategy based on dual detection of fronthaul bandwidth and spectrum units is introduced, and the instantaneous load of the system is controlled through dynamic migration batch division, effectively preventing chain overload and signaling storms caused by migration. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only for the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0020] Figure 1 Schematic diagram of a method flow in an embodiment of the present invention; Figure 2 Schematic diagram of dual-channel joint decision-making in an embodiment of the present invention. DETAILED DESCRIPTION

[0021] The present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. Those skilled in the art may also implement some known technologies in other alternative ways. The accompanying drawings are only for describing the embodiments in more detail and are not intended to limit the present invention in any specific way.

[0022] like Figure 1-Figure 2 As shown in FIG, the interference coordination and load balancing method between cells in dense heterogeneous networks based on graph neural networks includes the following steps: S1. Constructing a heterogeneous network graph with enhanced business features: User terminals are regarded as first-class nodes; base stations are regarded as second-class nodes; heterogeneous edges are constructed based on user-base station association relationships and inter-base station interference relationships, including user-base station association edges and inter-base station interference edges; and a heterogeneous network graph that integrates the spatiotemporal characteristics of services is output.

[0023] S1 includes the user node definition: the user terminal in the active state is regarded as the first type of node, and its node attributes include: ① Service type: The service type of each user terminal is selected from the set {eMBB, URLLC, mMTC} and represented by a standardized identifier.

[0024] ② Position coordinates: the user's real-time position in the three-dimensional Cartesian coordinate system ; The location update period is no more than 100ms, which is used to support spatial interference mapping and topology change detection.

[0025] ③ Historical traffic fluctuation coefficient: The fluctuation factor in the form of standard deviation is calculated based on the traffic data of the past N time slots: ;in, Indicates the current time slot The historical flow fluctuation coefficient, is the length of the sliding time window, is the historical time slot index within the time window, is the current time slot number, For the The user traffic value of each time slot is Indicates the past The average flow value of a time slot is defined as: .

[0026] S1 includes the base station node definition: macro base stations, micro base stations, and pico base stations within the target area are all considered as second-class nodes. Their node attributes include: ① Base station height: The vertical height of the base station antenna phase center relative to the sea level, denoted as , used for fine-grained modeling of 3D path loss, Indicates base station The vertical height of the antenna phase center relative to sea level.

[0027] ② Remaining forward transmission bandwidth: The transmission capacity of the currently unoccupied forward transmission link, in Gbps, recorded as , indicating a base station The current remaining fronthaul link bandwidth.

[0028] ③ Available spectrum unit set: The set of physical resource blocks that can be dynamically scheduled by the base station, defined as: ;in, Indicates base station The set of currently available spectrum units, Indicates base station No. Available spectrum units, Indicates base station The number of currently available spectrum units (i.e., the set base number).

[0029] The construction of heterogeneous edges in S1 is divided into the following two categories: User-base station association edge (directed edge): A directed edge is established between user u and its home base station b. Its edge weight is defined as: ;in: represents the edge weight between user u and base station b, Represents a user The type of business, represents the real-time signal-to-noise ratio between user u and base station b, Represents a user The service type priority weight mapping function takes the value {eMBB, URLLC, mMTC}: eMBB → 0.8: Enhanced Mobile Broadband (eMBB) primarily serves high-bandwidth services such as HD video and AR / VR. It requires high speeds but has a high tolerance for latency. Therefore, it is assigned a medium priority weight of 0.8.

[0030] URLLC → 1.0: Ultra-Reliable Low Latency Communication (URLLC), which is used for mission-critical services such as telemedicine and autonomous driving. These services are extremely sensitive to latency and reliability, and are therefore given the highest priority of 1.0.

[0031] mMTC → 0.6: Massive Machine Type Communication (mMTC), such as IoT sensor networks, has low traffic volume and is not sensitive to latency. It is assigned the lowest priority weight of 0.6.

[0032] ② Interference edge between base stations (undirected edge): between any two base stations When the overlap of their coverage areas exceeds 10%, an undirected edge is established between them, and the edge weight is defined as: ;in, Respectively The set of users covered, represents the coverage overlap, which is defined as: ; Indicates that the base station The set of users covered, Indicates that the base station The set of users covered, Indicates that the base station and The set of users covered at the same time (intersection), Indicates that the base station or The set of users covered (union), It is the cardinality operator of the set, indicating the number of elements in the set; represents the spectrum coincidence index, which is defined as: ; Indicates base station The set of available spectrum units, Indicates base station The set of available spectrum units, Indicates base station and The intersection of spectrum units (commonly available spectrum resources), Indicates base station and The union of the spectrum units of (all spectrum resources involved by both), Represents the cardinality operator for a set, indicating the number of elements in the set.

[0033] Combining the above node and edge information, the real-time output heterogeneous network graph structure is as follows: ;in, Represents the overall structure of the heterogeneous network graph, Represents the user node set (i.e., all active user terminals), represents a set of base station nodes (including macro base stations, micro base stations, and pico base stations), represents the set of associated edges from the user to the base station, represents the set of interference edges between base stations, Represents the user node attribute matrix, where rows correspond to user nodes and columns are business types, three-dimensional coordinates, historical traffic fluctuation coefficients, etc. Represents the base station node attribute matrix, where rows correspond to base station nodes and columns are base station height, remaining fronthaul bandwidth, available spectrum set, etc. represents the edge weight matrix, which contains all and The numerical representation of , used for graph neural network calculations.

[0034] S2. Dual-channel graph neural network joint decision-making: The heterogeneous network graph is input into a dual-channel graph neural network model, including an interference coordination channel and a load balancing channel, wherein: Interference coordination channel: Through the spectrum perception graph convolution layer, it generates the spectrum efficiency gain distribution matrix of each base station, indicating the theoretical interference reduction range of each spectrum unit; Load balancing channel: Generates a load pressure transmission path through the service migration graph attention layer, identifies the migration priority of the users to be migrated and the target base station sequence.

[0035] The dual-channel graph neural network joint decision in S2 takes the heterogeneous network graph output by S1 as input, constructs a dual-channel graph neural network model for interference coordination and load balancing, extracts node status through multi-layer graph convolution and attention mechanism, and performs joint resource reconstruction decision.

[0036] S21, interference coordination channel processing flow: S211, spectrum perception graph convolution layer operation: for each base station node Perform the following neighborhood aggregation operations:

[0037] in, Indicates the Layer time base station The node embedding representation of Indicates the Layer time base station The node embedding representation of Indicates the Neighboring base stations The node embedding representation of Indicates the Layer Time User The node embedding representation of Indicates the Layer spectrum-aware convolution weight matrix, Indicates base station The set of connected user nodes, Indicates that the base station The set of neighboring base stations with interference edges, represents the rectified linear unit activation function, represents the user-base station attention weight, that is, the user node aggregation weight, represents the edge weight between user u and base station b, Represents candidate users The edge weight between β and base station b, represents the spectrum coupling coefficient between base stations, that is, the aggregation weight of base station nodes, Indicates base station b and base station The interference edge weights between Indicates base station b and base station The size of the spectrum intersection between Represents the maximum value of two numbers and is used to normalize the intersection weight.

[0038] S212, generate a spectrum efficiency gain distribution matrix: for each base station Each spectrum unit , execute the following mapping: ; in, Indicates base station In spectrum unit The theoretical interference reduction on represents the spectral gain mapping weight matrix, represents the Sigmoid activation function, represents the base station embedding state output by the last layer of the graph neural network, represents the path loss compensation factor, Indicates the spectrum attribution indicator variable. If the spectrum unit belong If yes, the value is 1, otherwise it is 0. is the set of available spectrum units of base station b, For adjacent base stations The set of available spectrum units, Indicates base station and The interference edge weights between Indicates that all spectrum units Neighboring base stations Perform the maximum value operation, is the path loss compensation factor, which is obtained by inversely deducing the statistical mean of the path loss between base stations in the wireless propagation model. Its function is to balance the influence of the interference edge weight at different distance scales. It can be obtained by regressing the path loss between different base station pairs in historical communication data. In actual networks, the empirical value of γ is 0.3–0.7, and the initial setting is γ=0.5.

[0039] The final output spectrum efficiency gain matrix is: ,in ; represents the field of real numbers, is the number of base station nodes, is the total number of all available spectrum units (i.e., spectrum dimension), Representation matrix Base station b in spectrum unit The theoretical interference reduction range is the specific element value in the matrix.

[0040] The spectrum-aware graph convolutional layer operation mentioned above updates the features of the graph neural network of the base station node. The core is: first, the base station node retains information from its own historical features; then, it aggregates features from the user nodes connected to it. The aggregation weight is determined by the edge weight between the user and the base station (reflecting link quality and service priority); at the same time, it also aggregates information from other base station nodes that have an interference relationship with it. The aggregation weight combines the interference intensity and spectrum overlap to more realistically characterize the transmission path of spectrum interference.

[0041] The present invention adopts this design so that when each base station updates its status in the graph neural network, it can perceive both the business demand characteristics of its subordinate users and the potential interference impact from neighboring base stations; compared with traditional methods that only consider the physical location or fixed topology aggregation between base stations, this method can more dynamically reflect the actual situation of interference propagation and spectrum coupling, thereby improving the accuracy of interference identification and suppression.

[0042] The spectrum efficiency gain distribution matrix generates and outputs the theoretical interference reduction potential of each base station in each available spectrum unit. Specifically, the final state of the base station's graph neural network, information about the current spectrum availability, and the maximum interference strength between neighboring base stations on the spectrum are input. Through a linear mapping layer and activation function, a normalized spectrum gain score is output. All spectrum gain scores are combined into a matrix to guide subsequent spectrum allocation operations.

[0043] The present invention adopts this design to obtain a more fine-grained spectrum resource efficiency estimation by fusing three types of information: graph embedding, spectrum attribution, and interference intensity. The output gain matrix can be used to screen those spectrum units that are most worthy of priority allocation in an interference environment, realizing fine-grained regulation of spectrum resources for interference reduction. Compared with the scheme of spectrum allocation based only on global average interference or static threshold, this design can more fully reflect the current state of the network and realize dynamic and adaptive spectrum coordination capabilities.

[0044] S22, load balancing channel processing flow: S221, business migration graph attention layer operation: for each user node Calculate its migration utility score to the candidate base station: ; in, Represents a user To the base station The utility score of the migration, represents the migration score weight matrix, Represents the user and base station embedding status output by the last layer of the graph neural network, Indicates base station Current remaining forward transmission bandwidth, Represents a user The historical flow fluctuation coefficient, represents the spatial coordinates of the user and the base station, Represents distance in Euclidean space.

[0045] Calculate the migration priority based on the current base station load rate and user service type: ; in, is the priority mapping function of the user service type, is the current load rate of the base station to which the user is connected, A collection of high-load users to be processed.

[0046] The migration utility score calculates the cost-effectiveness of migrating to a candidate base station, integrating four key pieces of information: The graph embedding state of users and base stations: reflects the global behavioral characteristics of the two currently learned in the graph neural network; The remaining fronthaul bandwidth of the target base station: This measures whether the base station can handle the data needs of newly migrated users; User traffic fluctuation coefficient: indicates whether user services are stable. The smaller the fluctuation, the easier it is to migrate smoothly. Spatial distance between the user and the target base station: The shorter the migration path, the better the channel conditions and the lower the migration cost.

[0047] These features are concatenated and fed into a scoring function, which outputs a numerical score that quantifies whether the migration is "worth it."

[0048] The present invention no longer uses signal strength or static distance as the sole basis for judgment during the load balancing process, but introduces user traffic stability and target base station bandwidth constraints to avoid service quality deterioration after migration, and realizes multi-factor fusion and refined modeling of migration benefit evaluation, which is closer to the actual network operation situation.

[0049] After completing the migration score for each user, a priority function is introduced to determine which users should be migrated first and where to migrate to. The migration urgency of each user is calculated based on the load rate of the currently connected base station and the user's service type. A high priority means that the user should be diverted as soon as possible.

[0050] Path construction function: From all candidate target base stations, select the target combination with the highest score and the least number of hops to form an ordered target sequence, so that the migration is more in line with the requirements of resource conservation and network topology efficiency.

[0051] Path attenuation: This feature introduces a hop attenuation factor to suppress long-path migrations, preventing transmission delays and network oscillations caused by excessively long migration chains.

[0052] The present invention ensures that high-priority users in high-load base stations are processed first, improves the "pressure relief efficiency" of the migration strategy, avoids selecting only the nearest or most idle base station, and instead combines the loadability with the dynamic evaluation of the overall network structure to realize the load pressure gradient release mechanism under multi-hop path planning. This is a key improvement that distinguishes it from the "one-time on-site migration" in the existing solution.

[0053] S222, build load pressure conduction path: press Select the user set to be migrated in descending order ; for each Generate target base station sequence ( is the index variable for migrating users): ,in ;in, Represented as a user The target base station sequence constructed, Represents a user The candidate migration base station sequence, where M is the number of candidate base stations, represents the set of all optional target base stations, Indicates the distance from the current base station to the candidate base station The number of hops, is the path attenuation factor, ranging from [0.1, 1.0]. In most dense scenarios, it is recommended to be initially set to 0.3.

[0054] The final output path set: ;in, Represents a triplet record item (user, priority, target base station sequence) in the migration path set.

[0055] Spectral gain mapping weight matrix The interference reduction estimate used to map the base station hidden state characteristics and interference environment information into spectrum units is obtained as follows: (a) Constructing a training dataset: Collect a large number of real or simulated network status snapshots, including the spectrum allocation status of each base station, graph embedding features, spectrum ownership information, and interference indicators of neighboring base stations, and annotate the actual spectrum efficiency changes of each spectrum unit (as a supervision target).

[0056] (b) Initialize the model structure: Design the graph neural network backbone, and use the linear transformation module in the output layer to receive the splicing vector , and output the predicted value .

[0057] (c) Using regression loss function: The mean square error between the predicted value and the annotated true spectrum efficiency improvement value is used as the loss function, and backpropagation optimization is performed.

[0058] (d) After multiple rounds of training, the converged linear mapping weight is , which can be used to generate spectrum allocation strategies in the inference phase.

[0059] Migration score weight matrix It is used to map the joint feature vector between the user and the target base station into a migration utility score, which is obtained as follows: (a) Constructing a migration sample dataset: Based on historical network load change records, we annotate which user migration behaviors are successful and effective (i.e., load balancing is improved and QoS is met after migration). At the same time, we extract features such as graph embedding, bandwidth-to-fluctuation ratio, and spatial distance for each user-target base station pair.

[0060] (b) Constructing a scoring model: concatenate the above features into a unified vector as the input of the scoring function, and use a single-layer linear mapping structure as the output layer. Its parameters are .

[0061] (c) Training phase: Use supervised learning to perform binary cross-entropy loss using labels indicating whether the transfer was successful or not, or use reinforcement learning to optimize the scoring function parameters using the improvement in network load as a reward signal.

[0062] (d) After the final training converges, the weight matrix obtained Used to score the migration feasibility of any user-base station combination in online scenarios and assist in load balancing path planning.

[0063] S3. Perform spectrum-load joint reconstruction: performing cross-base station orthogonal allocation on spectrum units having gain values ​​higher than a preset gain threshold according to the spectrum efficiency gain distribution matrix; Along the load pressure transmission path, users are migrated in batches according to priority to the first available base station in the target base station sequence.

[0064] S3 performs spectrum-load joint reconstruction based on the joint output of the interference coordination channel and the load balancing channel, performs cross-base station orthogonal reconstruction of spectrum units and user bandwidth migration control, and realizes the coordinated optimization of network interference suppression and resource load balancing.

[0065] S31, orthogonal allocation of spectrum elements: S311, gain threshold screening, setting the preset gain threshold of the spectrum efficiency gain: ;in, is the spectrum efficiency gain threshold, Indicates the adjustable coefficient, reflecting the current interference intensity sensitivity, Gain distribution matrix This dynamic threshold is designed by setting an adaptive reference line based on overall spectrum efficiency to screen spectrum units with true interference reduction potential, avoiding over-allocation or omission of high-quality resources caused by fixed threshold settings. Specifically, the global mean of the gain distribution matrix is ​​used as a benchmark to dynamically reflect the current interference environment and spectrum efficiency level of the entire network. This is then multiplied by an adjustable coefficient to enhance screening flexibility and ensure that the selected spectrum units have scheduling value significantly above the average level.

[0066] For each base station ,Filter high-value spectrum unit collection: ; That is: all spectrum efficiency gain values ​​in base station b are greater than the threshold That is, from all spectrum units of base station b, select the frequency points whose gain scores exceed the set threshold for subsequent priority allocation, where is the base station in the spectrum efficiency gain matrix In spectrum unit The predicted value on .

[0067] S312, cross-base station orthogonal allocation, considers each base station as a node in the graph. When two base stations have the same high-gain spectrum unit, it indicates that they have potential interference conflict, so an edge is established between them, that is: Construct a spectrum conflict graph: ; Among them, the node set For all base stations, the edge set :When any two base stations satisfy: That is, when there is a conflict between high-gain spectrum units, an edge is established between them. A graph coloring algorithm is applied to the conflict graph, with each color representing an orthogonal spectrum resource group. Adjacent base stations cannot share the same color. Each color represents an independent spectrum resource group, equivalent to an "orthogonal channel set."

[0068] Finally, the base station group corresponding to each color Spectrum units are uniformly allocated to an orthogonal resource pool to avoid spectrum reuse conflicts in spatially adjacent areas.

[0069] S32, user migration execution: S321, Migration batch division: Group the users to be migrated In descending order of priority: ; satisfy: , that is, The migration priority of users is no lower than The priority of each user.

[0070] After the sorting is complete, not all users are migrated at once. Instead, the number of users that can be migrated simultaneously, i.e., the capacity of a migration "batch," is calculated based on the available fronthaul bandwidth resources of the current network. Specifically, the maximum batch size allowed is calculated by dividing the total available fronthaul bandwidth by the bandwidth required for a single user migration. The entire user set is then divided into several migration batches based on this size. This division avoids the simultaneous migration of a large number of users in a short period of time, which could cause signaling storms or instantaneous overload of the target base station, and adapts to the current resource status of the system; that is, setting the migration batch size: ;in, Indicates the total fronthaul bandwidth currently available on the network. represents the average forward bandwidth required for a single user migration, Indicates a floor operation.

[0071] Divide the sorted user set into Migration batches are executed sequentially. Indicates rounding up, that is, the total number of users Maximum per batch When grouping people, the number of batches required, even if the last group has insufficient people , but also counted separately.

[0072] S322, target base station selection, for the Each user in a migration batch , along the target base station sequence: ; Sequential detection: During user migration, a target base station is selected. A candidate base station sequence is generated in advance for each user to be migrated, representing possible migration destinations. Each base station in this sequence is checked in turn to determine whether it has sufficient resources to accommodate the user. If a base station meets two conditions, it is selected as the final migration target: Sufficient fronthaul bandwidth: The remaining fronthaul bandwidth of the base station must be greater than or equal to a certain threshold of user bandwidth requirements (usually used to ensure redundancy after migration to avoid immediate overload). Sufficient spectrum resource redundancy: The number of spectrum units available to the base station is no less than twice the spectrum required by users, which is used to ensure the flexibility and stability of resource scheduling.

[0073] Once the first base station that satisfies both conditions is found, the system selects it as the user's migration target and stops traversing subsequent candidates.

[0074] Specifically, it is to select the first base station that meets the following conditions : ; in, Indicates base station Current remaining forward transmission bandwidth, Represents a user Business bandwidth requirements, Indicates the bandwidth margin ratio threshold, with a value of 1.3, indicating that the remaining forward transmission bandwidth of the target base station should be at least 130% of the user service bandwidth requirement. Indicates base station The set of currently available spectrum units, Represents a user The number of spectrum units required. If the conditions are met, the base station is marked as the user's migration target.

[0075] S323, migration operation: Perform the following migration operations on the selected migration target base station: Allocate required spectrum units in its orthogonal resource pool for the migrating user; User context and data plane switching is completed through the X2 interface; Update the load rate, remaining bandwidth, and spectrum occupancy of the source and target base stations.

[0076] The present invention encompasses any alternatives, modifications, equivalents, and solutions that fall within the spirit and scope of the present invention. To provide a thorough understanding of the present invention, specific details are described in detail below in connection with the preferred embodiments of the present invention, but those skilled in the art will be able to fully understand the present invention without these detailed descriptions. Furthermore, to avoid unnecessary confusion regarding the essence of the present invention, well-known methods, processes, procedures, components, and circuits have not been described in detail.

[0077] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be regarded as within the scope of protection of the present invention.

Claims

1. A method for interference coordination and load balancing between cells in dense heterogeneous networks based on graph neural networks, characterized by: The following steps are involved: S1. Constructing a heterogeneous network graph with enhanced business features: The user terminal is regarded as the first type of node; the base station is regarded as the second type of node; based on the user-base station association relationship and the interference relationship between base stations, heterogeneous edges are constructed, wherein the heterogeneous edges include user-base station association edges and inter-base station interference edges; Output a heterogeneous network graph that integrates the spatiotemporal characteristics of services; S2. Dual-channel graph neural network joint decision-making: The heterogeneous network graph is input into a dual-channel graph neural network model, including an interference coordination channel and a load balancing channel, wherein: Interference coordination channel: Through the spectrum perception graph convolution layer, it generates the spectrum efficiency gain distribution matrix of each base station, indicating the theoretical interference reduction range of each spectrum unit; Load balancing channel: Generates a load pressure transmission path through the service migration graph attention layer, identifies the migration priority of the users to be migrated and the target base station sequence; S3. Perform spectrum-load joint reconstruction: performing cross-base station orthogonal allocation on spectrum units having gain values ​​higher than a preset gain threshold according to the spectrum efficiency gain distribution matrix; Along the load pressure transmission path, users are migrated in batches according to priority to the first available base station in the target base station sequence.

2. The method for interference coordination and load balancing between cells in dense heterogeneous networks based on graph neural networks according to claim 1 is characterized in that: The node attributes of the first type of nodes include service type, location coordinates, and a historical traffic fluctuation coefficient, where the historical traffic fluctuation coefficient is calculated based on a standard deviation of traffic over multiple time slots in the past.

3. The method for interference coordination and load balancing between cells in dense heterogeneous networks based on graph neural networks according to claim 2, characterized in that: The node attributes of the second type of nodes include base station height, remaining fronthaul bandwidth, and available spectrum units.

4. The method for interference coordination and load balancing between cells in dense heterogeneous networks based on graph neural networks according to claim 3 is characterized in that: The edge weight of the user-base station association edge is calculated based on the weighted coupling of the user service type and the link quality index. The QoS levels corresponding to different service types are converted into numerical weights through a preset service type priority mapping function. The weights are then combined with the real-time signal-to-noise ratio between the user and the home base station to obtain a link quality index oriented to service performance. The edge weight of the inter-base station interference edge is calculated based on the product of spatial overlap and spectrum overlap. The overlap of the coverage areas is quantified by calculating the ratio of the number of shared service users between the two base stations to the total number of covered users. The degree of sharing of spectrum resources is calculated by the intersection-over-combination ratio of the available spectrum sets of the two base stations. The edge weight of the inter-base station interference edge reflects the interference intensity caused by the coupling of spatial topology and spectrum resources. The final output heterogeneous network graph includes the user node set, the base station node set, the user-to-base station association edge set, the interference edge set between base stations, the user node attribute matrix, the base station node attribute matrix, and the edge weight matrix.

5. The method for interference coordination and load balancing between cells in dense heterogeneous networks based on graph neural networks according to claim 1, characterized in that: The interference coordination channel is used to model the spectrum interference relationship between base stations. It adopts the spectrum perception graph convolution mechanism to perform dual-source neighborhood feature aggregation on the base station node, and fuses the information from adjacent user nodes and adjacent base station nodes. The user node aggregation weight is generated based on the service perception attention function, and the base station node aggregation weight is calculated based on the spectrum overlap degree and the interference edge weight. After feature aggregation, the spectrum efficiency gain prediction value of each base station on each available spectrum unit is generated. , indicating a base station In spectrum unit The theoretical interference reduction amplitude on the spectrum efficiency gain is obtained, and the spectrum efficiency gain prediction value is combined with the base station hidden state, spectrum ownership identifier and the strongest interference link information, and is output through linear prediction and activation function; all spectrum efficiency gain prediction values ​​constitute a spectrum efficiency gain distribution matrix.

6. The method for interference coordination and load balancing between cells in dense heterogeneous networks based on graph neural networks according to claim 1, characterized in that: The load balancing channel models the feasibility and priority of user migration. Based on the graph attention mechanism, a migration utility scoring function is constructed between each user and the candidate target base station to obtain a migration utility score. The migration utility score integrates the graph embedding state of the user and the base station, the remaining forward transmission bandwidth of the target base station, the user traffic fluctuation coefficient and the spatial distance. Based on the load rate and service type priority of the base station to which the user is currently connected, the migration priority is calculated and the set of users to be migrated is screened.

7. The method for interference coordination and load balancing between cells in dense heterogeneous networks based on graph neural networks according to claim 6, characterized in that: The load balancing channel also includes generating an ordered target base station sequence for each user to be migrated based on the migration utility score and hop count attenuation mechanism, constructing a load pressure transmission path, and outputting a path set including user identification, migration priority and target base station sequence for load balancing scheduling.

8. The method for interference coordination and load balancing between cells in dense heterogeneous networks based on graph neural networks according to claim 1, characterized in that: The S3 includes spectrum unit screening, specifically including calculating the global average value of the spectrum efficiency gain matrix, and combining the adjustable coefficient to obtain the preset gain threshold under the current time slot, screening the spectrum efficiency gain value of each base station, and only retaining the spectrum units greater than the preset gain threshold to form a set of high-value spectrum units of the current base station as candidate resources for priority scheduling.

9. The method for interference coordination and load balancing between cells in dense heterogeneous networks based on graph neural networks according to claim 8, characterized in that: The cross-base station orthogonal allocation includes taking all base stations as nodes. If two base stations have overlapping frequency points in the high-gain spectrum set, a connection edge is established between them to form a spectrum conflict graph. A graph coloring operation is performed on the conflict graph, and all base stations are grouped. It is required that there is no spectrum conflict between the base stations in each group. The respective high-gain spectrum units in each group are uniformly assigned to the same orthogonal resource pool to ensure that adjacent base stations do not transmit at the same frequency point, thereby achieving interference avoidance and efficient reuse of spectrum resources.

10. The method for interference coordination and load balancing between cells in dense heterogeneous networks based on graph neural networks according to claim 9, characterized in that: Said S3 further comprises: Migration batch division: All users to be migrated are sorted from high to low according to migration priority, with priority given to users in high-service-level or high-load areas. The number of users that can be accommodated in each migration batch is calculated based on the available fronthaul bandwidth of the current network and the average bandwidth requirements of a single user. All sorted users are then divided into several batches, and migration is carried out gradually in batches. Target base station selection: For each user in each batch, the pre-generated target base station sequence is traversed, and the resource status of each candidate base station is checked in turn. If the remaining fronthaul bandwidth of a base station exceeds the set safety threshold of the user's bandwidth demand and the number of its available spectrum units is at least twice the user's demand, the current base station is considered to have sufficient carrying capacity and is immediately marked as the user's migration target. Migration operation: allocates corresponding spectrum resources to the user in the orthogonal spectrum pool of the target base station, and completes the transfer of user context and data plane switching through the interface.

Images