Multi-user MIMO Beam and Power Allocation Method Based on Graph Attention Expansion

Abstract

This invention discloses a multi-user MIMO beamforming and power allocation method based on graph attention expansion, comprising the following steps: Step 1: Constructing a multi-user interference channel MIMO system model and obtaining the channel matrix; Step 2: Expanding the WMMSE algorithm iteration process into a multi-layer neural network structure, and introducing a graph attention mechanism into the expanded network to model the interference relationship between users; Step 3: Obtaining the beamforming matrix through network iterative optimization, performing adaptive power allocation, and outputting the joint optimization result of beamforming and power. The beneficial effects of this invention are: This invention combines deep learning, attention mechanisms, algorithm expansion, and the WMMSE algorithm to construct a multi-user MIMO resource allocation scheme. This scheme can reduce the complexity of the neural network, enhance the feature representation capability of the neural network, and thus improve the system's real-time performance and spectral efficiency.

Description

Technical Field

[0001] This invention belongs to the field of wireless communication technology, specifically relating to a multi-user MIMO beam and power allocation method based on graph attention unfolding. Background Technology

[0002] With the continuous evolution of sixth-generation mobile communication systems (6G), future wireless networks will exhibit significant characteristics such as ultra-high connection density, ultra-large-scale collaboration, and high heterogeneity. Emerging application scenarios, such as immersive communication and massive machine-type communication, place unprecedented demands on network capacity and concurrency capabilities, inevitably leading to a complex network environment characterized by high user concurrency and highly coupled interference. As the user base continues to grow, interference caused by multi-user concurrency intensifies significantly. Even under relatively abundant bandwidth resources, system spectral efficiency may still decline significantly due to the inability of resource allocation and interference suppression mechanisms to adapt to the large-scale interference coupling structure. To improve spectrum utilization efficiency, systems typically employ spatial multiplexing or co-frequency concurrent transmission mechanisms, enabling multiple users to simultaneously transmit or receive data within the same frequency band at the same time. However, under shared spectrum conditions, signals from different transmitting nodes superimpose in space, causing the receiver to simultaneously receive unwanted signal components from other users in addition to the desired signal, thus creating multi-user interference. Beamforming is a key technology for suppressing interference and enhancing desired signals by utilizing spatial dimensions. Transmit power is an important resource for reducing interference and improving system spectral efficiency. Therefore, by rationally allocating beam and power resources, interference can be effectively mitigated and system spectral efficiency can be improved, thereby meeting the needs of wireless communication system development.

[0003] Traditional resource allocation methods based on optimization theory have been extensively studied in wireless communication systems. These methods typically rely on mathematical tools such as convex optimization, alternating optimization, or fractional programming, and can obtain optimal or near-optimal solutions under certain conditions, possessing a sound theoretical foundation and interpretability. However, with the continuous increase in the number of users and the scale of the system, the dimensionality of the optimization problem increases significantly, leading to high computational complexity. Furthermore, these methods usually require multiple iterations to obtain a convergent solution, making them difficult to meet the real-time processing requirements in dynamic environments with rapidly changing channels.

[0004] To overcome the high complexity and poor generalization ability of traditional optimization methods, deep learning methods have received widespread attention in recent years. Thanks to the powerful nonlinear modeling capabilities of deep neural networks, these methods can learn complex channel characteristics and inter-user interference relationships from data, and quickly obtain resource allocation results through a single forward propagation after model training, thus significantly reducing online computational complexity. However, purely data-driven deep learning methods typically rely on large amounts of training data, their model decision-making processes lack clear structural explanations, and their generalization ability in different network environments still needs further improvement.

[0005] To combine the structural advantages of traditional optimization methods with the high computational efficiency of deep learning methods, model-driven deep unfolded networks have gradually become a research hotspot in recent years. This type of method maps traditional iterative optimization algorithms to a neural network structure with a fixed number of layers, introducing learnable parameters while retaining the original optimization ideas. This improves computational efficiency while maintaining performance, providing a new research approach for resource allocation problems in complex wireless networks.

[0006] However, although model-driven deep unfolding networks have overcome some of the shortcomings of traditional optimization algorithms and deep learning to a certain extent, they still face some challenges and limitations in practical applications, specifically in the following aspects: (1) Feature extraction lacks interference selectivity, limiting the potential for improving spectral efficiency: In previous studies, traditional GNNs typically use fixed or uniform weights to fuse the features of neighboring nodes during information aggregation, failing to adaptively adjust the importance of different neighbors based on node features. In multi-user interference networks, due to the significant differences in interference intensity between different links, this mechanism struggles to accurately characterize the interference coupling relationship between users, thus limiting the model's ability to express complex interference structures and the precision of beam optimization.

[0007] (2) Rigid power allocation strategy and lack of energy efficiency optimization: Existing UWMMSE methods adopt a hard power constraint strategy, which only meets the power limit by "full power transmission" or "proportional contraction", without dynamic allocation based on instantaneous channel quality. This strategy ignores the differences in channel conditions among users: deep fading users transmit at full power, resulting in low energy efficiency, and strong interference users fail to reduce power to reduce network interference. At the same time, the single shared Lagrange multiplier μ cannot adapt to differentiated channels, and the optimization objective only focuses on spectral efficiency without explicitly considering energy efficiency, causing the network to tend to operate at high power. This power-rate disconnect optimization leads to severe degradation of energy efficiency, and the spectral efficiency does not reach the theoretical optimum.

[0008] A prior art method and system for wireless power allocation based on an edge-aggregating graph attention network (publication number: CN118488554A) is disclosed. This method introduces the concept of an edge-aggregating graph attention network, abstracting the wireless ad hoc network into a graph neural network structure. By embedding edge-aggregating graph attention in the unfolded WMMSE iterative layer to learn the slope and offset of auxiliary parameters, it achieves power allocation optimization and channel capacity improvement in multi-user wireless communication networks. However, this technical solution does not deeply integrate the multi-head graph attention mechanism with the WMMSE algorithm unfolding process to generate weight correction terms and introduce a complex channel masking mechanism. Therefore, it cannot achieve joint beamforming and power allocation optimization for multi-user MIMO systems, nor can it perform adaptive water-filling power allocation based on the signal-to-interference-plus-noise ratio (SINR) to simultaneously maximize system spectral efficiency and energy efficiency. Summary of the Invention

[0009] The main objective of this invention is to overcome the shortcomings of the prior art and propose a joint beam and power resource allocation scheme for multi-user MIMO systems based on attention unfolding. The system constructed by this scheme can reduce system complexity and significantly enhance the model's ability to express channel characteristics, ultimately improving spectral efficiency and energy efficiency.

[0010] The technical solution adopted by this invention to achieve the above objectives is: a multi-user MIMO beam and power allocation method based on graph attention unfolding, comprising the following steps: Step 1: Construct a multi-user interference channel MIMO system model and obtain the channel matrix; Step 2: Expand the WMMSE algorithm iteration process into a multi-layer neural network structure, and introduce a graph attention mechanism into the expanded network to model the interference relationship between users; Step 3: Obtain the beamforming matrix through network iterative optimization, perform adaptive power allocation, and output the joint optimization results of beam and power.

[0011] Preferably, the system in step 1 uses a geometric random channel model to generate a wireless channel matrix to describe the wireless propagation characteristics between the transmitter and receiver. The channel matrix from user j to user i can be represented as:

[0012] in, This represents the path loss coefficient, which is related to the distance between the transmitter and receiver; This represents the small-scale fading components, which typically follow a complex Gaussian distribution. By randomly generating user location coordinates and calculating path loss based on distance, a complete set of channel matrices for the system is obtained, serving as the data source for subsequent algorithm training and testing.

[0013] Preferably, the received signal vector of the system signal model in step 1 is:

[0014] in, This indicates that transmitter i is used to send signals to its assigned receiver r(i). Beamformer This indicates that it follows a complex Gaussian distribution. (0, Additive noise, of which This represents the variance of additive noise in the system. Let r(i) represent the communication channel between transmitter i and its assigned receiver r(i), and for all , This represents the interference between r(i) and all other transmitters j. This represents the beamforming matrix of the j-th user, which contains both spatial beam direction information and implicitly includes the transmit power.

[0015] Preferably, a graph attention mechanism is further introduced into the WMMSE unfolded network to enhance the model's ability to model multi-user interference relationships. In step 2, each user is treated as a graph node, and the channel interference between users is treated as graph edges. The query vector, key vector, and value vector corresponding to the user features are obtained through linear mapping. The attention weights between nodes are calculated and normalized. Based on these weights, the user features are weighted and summed to update the node features, thus realizing the modeling of multi-user interference relationships. Specifically, each user in the system is considered as a node in the graph structure, and the channel interference relationship between different users is considered as the connection edge between nodes. The input user features are first processed through linear mapping to obtain the query vector, key vector, and value vector, and then the attention weights between different nodes are calculated. By normalizing the attention weights, the degree of interference influence between different users is obtained, and the feature vectors are weighted and summed using these weights to obtain the updated node feature representation.

[0016] Preferably, in step 2, to further enhance the model's expressive power, the network employs a multi-head attention mechanism to model different interference modes in parallel. The graph attention mechanism can adaptively learn the importance weights of different neighboring nodes based on node features. Since interference intensity is affected by multiple factors such as channel state, beam direction, and receiving filter, the receiving filter u and transmitting beam v in the current unfolded layer are used as node features to characterize the user's instantaneous communication state. Specifically, in the... In a layer-expanded network, the node feature vector of the i-th user is defined as:

[0017] This feature vector contains information from both the receiver and the transmitter, reflecting the user's transmit and receive status in the current iteration layer, where || represents vector concatenation.

[0018] Preferably, by compressing the features through a learnable linear mapping and projecting them into a low-dimensional latent space using the ReLU activation function

[24] , the complexity of graph attention computation can be reduced while preserving key transmission and reception and interference information, and overfitting of the original physical model can be avoided. The specific expression of the node features after linear mapping and ReLU activation is as follows:

[0019] in, , These are trainable parameters.

[0020] Preferably, in step 2, physical channel topology information is introduced, and a masking mechanism is introduced based on the channel coefficients: when there is no effective channel between transmitting user j and receiving user i, their corresponding attention score is forcibly suppressed. Finally, the attention weights are obtained through Softmax normalization, and the graph attention coefficients and normalized weights are calculated as follows:

[0021] in, This indicates taking the real part of a complex number.

[0022] Preferably, step 2 involves injecting the weight correction term of the graph attention output into the WMMSE weight matrix update process to complete the beam optimization iteration.

[0023] Preferably, in the online testing phase of step 3, after inputting the channel matrix H, the neural network updates the channel features layer by layer through the WMMSE expansion network, and models the interference relationship between users by combining the graph attention mechanism. After multi-layer network calculation, the beamforming matrix V of each user is output. This beamforming matrix describes the spatial processing method of each user's signal at the transmitter, so as to achieve effective suppression of multi-user interference.

[0024] Preferably, in step 3, after obtaining the beamforming matrix, a power allocation strategy is introduced. Under the condition of fixed beam direction, the instantaneous SINR of each user is calculated according to the current beam matrix, and the total available power is normalized and allocated according to the relative size of SINR. Under the condition of satisfying the total power constraint of the system, the transmit power of each user is dynamically adjusted.

[0025] (1) A large-scale MIMO communication system model under multi-user interference channels is constructed. Wireless channel data is generated through geometric channel modeling, taking into account the interference relationship between users and the impact of path loss and small-scale fading on channel characteristics. In this system model, the channel matrix is ​​used as input. By analyzing the signal transmission relationship between the transmitter and receiver, a multi-user signal transmission model is established. Under the constraint of total transmit power, the optimization objective is to maximize the system spectral rate and energy efficiency, thus providing a theoretical basis for subsequent algorithm design.

[0026] (2) Based on the above system model, a FWMMSE_GAT neural network model based on the idea of ​​deep expansion is constructed. The iterative process of the traditional weighted minimum mean square error (WMMSE) algorithm is expanded into a multi-layer neural network structure, so that each layer corresponds to one iteration update of the WMMSE algorithm, thereby improving the computational efficiency while maintaining the interpretability of the algorithm.

[0027] (3) Introduce graph attention mechanism into the unfolded network. By constructing a multi-head graph attention network, the interference relationship between users is modeled, so that the network can adaptively learn the interference weights between different users, thereby improving the model's ability to express complex wireless network topology and enhancing the system's adaptability in dynamic channel environment.

[0028] (4) In the final stage of the network structure, a power allocation mechanism based on signal-to-noise ratio (SINR) is introduced. The transmission power of each user is dynamically adjusted through the water-filling algorithm, so that the system can achieve better resource allocation effect under the premise of meeting the total power constraint.

[0029] Using the above methods, an optimized beam and power allocation scheme is finally obtained, enabling efficient resource allocation and interference management in a multi-user MIMO system.

[0030] The beneficial effects of this invention are as follows: This invention combines deep learning, attention mechanisms, algorithm expansion, and the WMMSE algorithm to construct a multi-user MIMO resource allocation scheme. This scheme can reduce the complexity of neural networks, enhance the feature representation ability of neural networks, and thus improve the real-time performance and spectral efficiency of the system. Furthermore, this invention integrates the WMMSE expansion algorithm and attention mechanisms, effectively improving the algorithm's adaptive capability, enabling the system to maintain high performance in complex channel environments and significantly reducing system processing time. Attached Figure Description

[0031] Figure 1 This is a schematic diagram of the neural network structure of the FWMMSE_GAT system. Detailed Implementation

[0032] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0033] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The following description of at least one exemplary embodiment is merely illustrative and is in no way intended to limit the present invention or its application or use. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0034] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0035] Example 1: A multi-user MIMO beam and power allocation method based on graph attention unfolding, comprising the following steps: Step 1: Construct a multi-user interference channel MIMO system model and obtain the channel matrix; Step 2: Expand the WMMSE algorithm iteration process into a multi-layer neural network structure, and introduce a graph attention mechanism into the expanded network to model the interference relationship between users; Step 3: Obtain the beamforming matrix through network iterative optimization, perform adaptive power allocation, and output the joint optimization results of beam and power.

[0036] Example 2: This example is based on the previous example, and the specific implementation further illustrates the present invention. A multi-user MIMO beam and power allocation method based on graph attention unfolding includes the following steps: Step 1: Establish the framework of a large-scale MIMO communication system under multi-user interference channels, defining the basic environment and constraints for resource allocation. In this system, multiple transmitters and receivers communicate simultaneously, and mutual interference exists between users. By constructing a channel model, the system's channel state information is obtained and used as input data for subsequent neural network algorithms, thus providing a foundation for resource allocation optimization.

[0037] (1) Channel Modeling: The system uses a geometric random channel model to generate a wireless channel matrix to describe the wireless propagation characteristics between the transmitter and receiver. Assume there are M users in the system, and each user's transmitter and receiver are equipped with T and R antennas respectively. Then the channel matrix from user j to user i can be expressed as:

[0038] in, This represents the path loss coefficient, which is related to the distance between the transmitter and receiver; This represents the small-scale fading components, which typically follow a complex Gaussian distribution. By randomly generating user location coordinates and calculating path loss based on distance, a complete set of channel matrices for the system is obtained, serving as the data source for subsequent algorithm training and testing.

[0039] System signal model: Let the transmitted signal vector be... If the beamforming matrix is ​​V, then the transmitted signal can be expressed as:

[0040] The received signal vector at the receiving end is:

[0041] in, This indicates that transmitter i is used to send signals to its assigned receiver r(i). Beamformer This indicates that it follows a complex Gaussian distribution. (0, Additive noise, of which This represents the variance of additive noise in the system. Let r(i) represent the communication channel between transmitter i and its assigned receiver r(i), and for all , This represents the interference between r(i) and all other transmitters j. This represents the beamforming matrix of the j-th user, which contains both spatial beam direction information and implicitly includes the transmit power.

[0042] (3) Problem Modeling: Assuming that the signals of different users are independent of each other and are not affected by receiver noise, the SINR of receiver r(i) is expressed as: Where H denotes the conjugate transpose, I is the identity matrix, and the channel state information tensor is defined. Make And define the transmitter beamformer tensor. Make For each user i, assuming complete knowledge of the CSI matrix, in a multi-user interference network, multiple users share the same frequency band for communication. The system spectral efficiency is defined as the total transmission rate per unit bandwidth. Therefore, the goal of this paper is to maximize the system spectral efficiency while satisfying the power constraints of each transmitter, i.e.:

[0043] Step 2: This scheme constructs a neural network model that integrates deep unrolling and graph attention mechanisms. Using the channel matrix as input, it adaptively generates the beamforming matrix for a multi-user system, thereby achieving efficient optimization of the wireless resource allocation problem. This network expands the iterative process of the traditional WMMSE algorithm into a multi-layer neural network structure and combines it with a graph attention mechanism to model the interference relationships between users, enabling the model to achieve efficient resource allocation in complex interference environments.

[0044] First, the system's channel matrix is ​​used as the network input through the input layer. Assume there are M users in the system, each user's transmitter and receiver are equipped with T and R antennas respectively. Then the system channel matrix can be expressed as:

[0045] The input channel matrix is ​​first preprocessed before being fed into the neural network, providing a foundation for subsequent feature extraction and resource optimization.

[0046] In the network structure, the iterative update process of the traditional WMMSE algorithm is expanded into a multi-layer network structure, with each layer corresponding to one iteration of the WMMSE algorithm. By updating the receiver filter matrix, weight matrix, and beamforming matrix layer by layer, the network can gradually approach the optimal resource allocation strategy. In each layer, the update coefficients in the traditional algorithm are learned through trainable parameters, thereby enhancing the model's adaptability in complex channel environments.

[0047] A graph attention mechanism is further introduced into the WMMSE unfolded network to enhance the model's ability to model multi-user interference relationships. Specifically, each user in the system is regarded as a node in a graph structure, and the channel interference relationship between different users is regarded as the connection edge between nodes. The input user features are first obtained by linear mapping to obtain query vector, key vector, and value vector, and then the attention weights between different nodes are calculated. By normalizing the attention weights, the degree of interference influence between different users is obtained, and the feature vectors are weighted and summed using this weight to obtain the updated node feature representation.

[0048] To further enhance the model's expressive power, the network employs a multi-head attention mechanism to model different interference modes in parallel. The graph attention mechanism can adaptively learn the importance weights of different neighboring nodes based on node features. Since interference intensity is affected by multiple factors such as channel state, beam direction, and receiver filter, this paper uses the receiver filter u and transmit beam v in the current unfolded layer as node features to characterize the user's instantaneous communication state. Specifically, in the... In a layer-expanded network, the node feature vector of the i-th user is defined as:

[0049] The feature vector contains information from both the receiver and transmitter, reflecting the user's transmit and receive status in the current iteration layer, where || represents vector concatenation. Then, the features are compressed through a learnable linear mapping and projected into a low-dimensional latent space through the ReLU activation function

[24] . This can reduce the computational complexity of graph attention while preserving key transmit and receive and interference information, and avoid overfitting the original physical model. The specific expression is as follows:

[0050] in, , These are trainable parameters. In graph attention networks, the interference relationships between users are characterized by attention coefficients. For any two user nodes i and j, their attention coefficients reflect the degree to which user i contributes to the weight correction of user j. In the... In layer expansion, the attention coefficient of the k-th attention head is defined as:

[0051] in, , , These are trainable parameters. To incorporate physical channel topology information, this paper introduces a masking mechanism based on channel coefficients: when there is no effective channel between transmitting user j and receiving user i, their corresponding attention scores are forcibly suppressed. Finally, the attention weights are obtained through Softmax normalization. in, This represents taking the real part of a complex number. This normalization process ensures that the sum of the attention weights for all neighboring nodes of user i is 1, thus forming a stable weighted aggregation mechanism. Then, the multi-head GAT uses multiple attention heads to compute the attention score of each head in parallel. Finally, the aggregated feature of a node is the weighted sum of the features of its neighbors, as shown in the formula:

[0052] To improve the model's expressive power and training stability, this paper employs a multi-head attention mechanism and introduces a residual connection structure. Specifically, the... The outputs of multiple attention heads in the layer are concatenated, then projected back into the low-dimensional feature space through a linear mapping, and finally added to the input features as residuals.

[0053] This operation enhances the model's ability to model complex perturbation relationships while effectively mitigating the gradient vanishing problem in deep unfolded structures.

[0054] After completing the graph attention feature aggregation, a fully connected layer is finally used as the output layer to map the node features, generating compensation terms to correct the traditional WMMSE weight matrix. In the layer, the weight adjustment term for the i-th user is defined as:

[0055] Finally, the first The layer weight matrix update form is as follows:

[0056] In this way, the global interference information learned by the graph attention network is explicitly injected into the WMMSE weight update process, enabling the model to adaptively optimize the update strategy according to the specific network topology, thereby improving the system's performance in complex interference environments.

[0057] During the offline training phase, batch training is employed to optimize network parameters. For each training batch, model parameters are updated by calculating system performance metrics and constructing a loss function. This scheme uses system spectral efficiency as the primary optimization objective, optimizing resource allocation strategies by maximizing system spectral efficiency.

[0058] Step 3: After the neural network completes offline training, all parameters in the network (including weight parameters and bias parameters) are determined and saved as the trained model. During actual communication system operation, network training is no longer performed; instead, the trained model is directly used for online inference. First, the real-time acquired channel matrix is ​​input into the neural network model, and the beamforming matrix of the system is calculated through forward propagation, thereby achieving rapid optimization of resource allocation in multi-user communication systems.

[0059] During the online testing phase, after inputting the channel matrix H, the neural network updates the channel features layer by layer through WMMSE expansion and models the interference relationships between users using a graph attention mechanism. After multi-layer network computation, the final output is the beamforming matrix V for each user. This beamforming matrix describes the spatial processing method of each user's signal at the transmitter, thereby achieving effective suppression of multi-user interference.

[0060] To further improve system resource utilization efficiency, a power allocation strategy is introduced after obtaining the beamforming matrix to dynamically adjust the transmit power of each user. Under the condition of satisfying the total system power constraint, the power is reasonably allocated according to the user channel quality, thereby improving the overall system performance. The complete process can be summarized as follows: Under the condition of fixed beam direction, the module first calculates the instantaneous SINR of each user based on the current beam matrix, and then normalizes and allocates the total available power according to their relative magnitude.

[0061] Furthermore, to comprehensively evaluate the algorithm's performance, system energy efficiency and user fairness metrics were also calculated. These metrics allow for a comprehensive evaluation of the algorithm from multiple dimensions, including spectral efficiency, energy efficiency, and user fairness.

[0062] Through the above online inference process, the trained neural network can quickly generate optimized beamforming strategies based on the real-time channel environment, thereby achieving efficient resource allocation and interference management in complex wireless environments.

Claims

1. A multi-user MIMO beam and power allocation method based on graph attention unfolding, characterized in that, Includes the following steps: Step 1: Construct a multi-user interference channel MIMO system model and obtain the channel matrix; Step 2: Expand the WMMSE algorithm iteration process into a multi-layer neural network structure, and introduce a graph attention mechanism into the expanded network to model the interference relationship between users; Step 3: Obtain the beamforming matrix through network iterative optimization, perform adaptive power allocation, and output the joint optimization results of beam and power.

2. The method according to claim 1, characterized in that, The system in step 1 uses a geometric random channel model to generate a wireless channel matrix. The channel matrix from user j to user i is represented as follows: in, Indicates the path loss coefficient; This represents the small-scale fading component.

3. The method according to claim 1, characterized in that, The received signal vector of the system signal model in step 1 is: in, This indicates that transmitter i is used to send signals to its assigned receiver r(i). Beamformer This indicates that it follows a complex Gaussian distribution. (0, Additive noise, of which This represents the variance of additive noise in the system. Let r(i) represent the communication channel between transmitter i and its assigned receiver r(i), and for all , This represents the interference between r(i) and all other transmitters j. Let represent the beamforming matrix for the j-th user.

4. The method according to claim 1, characterized in that, In step 2, each user is treated as a graph node and the channel interference between users is treated as a graph edge. The query vector, key vector and value vector corresponding to the user features are obtained through linear mapping. The attention weights between nodes are calculated and normalized. Based on the weights, the user features are weighted and summed to update the node features, thereby realizing the modeling of multi-user interference relationships.

5. The method according to claim 1, characterized in that, In step 2, the receiving filter u and the transmitting beam v in the current unfolded layer are used as node features to characterize the user's instantaneous communication state. In a layer-expanded network, the node feature vector of the i-th user is defined as: Here, || represents vector concatenation.

6. The method according to claim 5, characterized in that, The specific expression for the node features after linear mapping and ReLU activation is as follows: in, , These are trainable parameters.

7. The method according to claim 6, characterized in that, In step 2, when there is no effective channel between transmitting user j and receiving user i, the graph attention coefficient and normalized weight are calculated as follows: in, This indicates taking the real part of a complex number.

8. The method according to claim 1, characterized in that, Step 2 involves injecting the weight correction term from the graph attention output into the WMMSE weight matrix update process to complete the beam optimization iteration.

9. The method according to claim 1, characterized in that, In the online testing phase of step 3, after inputting the channel matrix H, the neural network updates the channel features layer by layer through the WMMSE expansion network, and models the interference relationship between users by combining the graph attention mechanism. After multi-layer network calculation, the beamforming matrix V of each user is output.

10. The method according to claim 1, characterized in that, In step 3, after obtaining the beamforming matrix, a power allocation strategy is introduced. Under the condition of fixed beam direction, the instantaneous SINR of each user is calculated according to the current beam matrix, and the total available power is normalized and allocated according to the relative size of SINR. Under the condition of satisfying the total power constraint of the system, the transmit power of each user is dynamically adjusted.

Images