Joint Channel Prediction and Beamforming Optimization Method Based on Fourier Graphical Neural Network

Abstract

This invention provides a joint channel prediction and beamforming optimization method based on Fourier graph neural networks. It estimates L consecutive historical channel state information timestamps based on signals received from a base station, and inputs this information into a target network trained using a hybrid training method to obtain a downlink beamforming matrix. The target network includes a channel prediction module, a power learning module, and a beamforming recovery module. The channel prediction module predicts the channel state information at any future time based on the L consecutive historical channel state information timestamps. The power learning module obtains the power characteristics of the future time based on the predicted channel state information. The beamforming recovery module obtains the downlink beamforming matrix based on the future channel state information, the future power characteristics, and the optimal beamforming solution structure, optimizing channel timeout and beamforming to maximize system performance and rate.

Description

Technical Field

[0001] This invention relates to the field of communication technology, and specifically to a joint channel prediction and beamforming optimization method based on Fourier graph neural networks. Background Technology

[0002] Beamforming technology offers advantages in signal focusing and interference suppression, demonstrating superior performance in wireless communication. This performance largely depends on the accuracy and real-time nature of Channel State Information (CSI). The more accurate the channel, the more effective the beamforming. However, due to the dynamic nature of wireless channels and the feedback delay of CSI, the CSI obtained at the base station is severely outdated, significantly reducing the beamforming effect and consequently causing a severe degradation in system performance. Furthermore, traditional algorithms that directly solve for the beamforming matrix using CSI (such as the Weighted Minimum Mean Squared Error (WMMSE) algorithm) typically require multiple iterations to find a local optimum. This can lead to high computational complexity and long convergence times, especially in high-dimensional problems where a large number of iterations are needed to converge to the optimal solution. To address these issues, joint channel prediction and beamforming design become crucial for optimizing communication performance.

[0003] Traditional channel prediction methods typically employ models such as Recurrent Neural Networks (RNNs), Long Short-Term Memory (LSTM) networks, and Gated Recurrent Units (GRUs) to study the temporal correlations between channels. However, for multi-user multiple-input single-output (MISO) systems, each user's channel response often exhibits significant spatial correlations, and mere temporal correlation may be insufficient for accurate channel prediction. Some emerging methods use graph neural networks to address channel staleness, but these methods either heavily rely on predefined graph structures to specify spatial correlations, failing to capture temporally changing spatial correlations, or they stack graph networks (such as Graph Convolutional Networks (GCNs) and Graph Attention Networks (GATs)) and temporal networks (such as LSTMs and GRUs) to separately capture spatial dynamics and temporal dependencies. This separate modeling of spatiotemporal features contradicts the uniformity of spatiotemporal correlations in the real world, severely impacting prediction performance. In addition, there are relatively few studies to date that simultaneously consider channel prediction and beamforming design.

[0004] Channel prediction methods based on Spectral-Temporal Graph Neural Networks (StemGNNs) utilize spectral sequence units and graph convolutions to capture the spatiotemporal correlation of the channel in the frequency domain. However, this invention learns temporal and spatial characteristics separately, a concept that contradicts the theory of unified spatiotemporal dynamics in the real world, severely impacting prediction performance. Furthermore, this technique does not consider the optimization of beamforming technology. Another existing method is a joint learning framework based on LSTM to address the beamforming problem, incorporating outdated CSI (Continuous Signal Processing). However, due to the limitations of LSTM, its performance needs improvement. Summary of the Invention

[0005] To address the aforementioned problems in the prior art, the present invention provides a method, apparatus, and device for decoding ordered statistics based on a confidence propagation list, specifically including:

[0006] In a first aspect, the present invention provides a joint channel prediction and beamforming optimization method based on a Fourier graph neural network, comprising:

[0007] Based on the signals received by the base station, L consecutive historical channel state information with timestamps are estimated. The signals received by the base station are signals sent by the user to the base station based on known pilot signals.

[0008] L consecutive historical channel state information with timestamps are input into the target network to obtain the downlink beamforming matrix through a hybrid training method.

[0009] The target network includes a channel prediction module, a power learning module, and a beamforming recovery module.

[0010] The channel prediction module is used to predict the channel state information at any future time based on L consecutive historical channel state information timestamps.

[0011] The power learning module is used to obtain the power characteristics of future time moments based on the predicted channel state information.

[0012] The beamforming recovery module is used to obtain the downlink beamforming matrix based on the channel state information, power characteristics, and optimal beamforming solution structure at future time moments.

[0013] Secondly, the present invention also provides a joint channel prediction and beamforming optimization device based on a Fourier graph neural network, comprising:

[0014] The acquisition module is used to estimate L consecutive timestamps of historical channel state information based on the signals received by the base station. The signals received by the base station are signals sent by the user to the base station based on known pilot signals.

[0015] The processing module is used to input the historical channel state information with L consecutive timestamps into the target network trained by the hybrid training method to obtain the downlink beamforming matrix;

[0016] The target network includes a channel prediction module, a power learning module, and a beamforming recovery module.

[0017] The channel prediction module is used to predict the channel state information at any future time based on L consecutive historical channel state information timestamps.

[0018] The power learning module is used to obtain the power characteristics of future time moments based on the predicted channel state information.

[0019] The beamforming recovery module is used to obtain the downlink beamforming matrix based on the channel state information, power characteristics, and optimal beamforming solution structure at future time moments.

[0020] Thirdly, the present invention also provides an electronic device, including a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus;

[0021] Memory, used to store computer programs;

[0022] The processor, when executing a program stored in memory, implements any of the methods provided in the first aspect.

[0023] The beneficial effects of this invention are:

[0024] This invention provides a joint channel prediction and beamforming optimization method based on Fourier graph neural networks. It estimates L consecutive historical channel state information at L timestamps based on signals received from the base station. This L consecutive historical channel state information is then input into a target network trained using a hybrid training method to obtain a downlink beamforming matrix. The target network includes a channel prediction module, a power learning module, and a beamforming recovery module. The channel prediction module predicts the channel state information at any future time based on the L consecutive historical channel state information. The power learning module obtains the power characteristics of the future time based on the predicted channel state information. The beamforming recovery module obtains the downlink beamforming matrix based on the future channel state information, the future power characteristics, and the optimal beamforming solution structure. This method considers the spatial correlation in multi-user, multi-antenna systems from a spatiotemporally dynamic perspective. Furthermore, it optimizes beamforming while addressing the channel timeout problem, reducing computational complexity, accelerating convergence, and achieving the goal of maximizing system performance and data rate.

[0025] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0026] Figure 1 A flowchart illustrating a joint channel prediction and beamforming optimization method based on a Fourier graph neural network provided by this invention.

[0027] Figure 2 A schematic diagram of the architecture of a target network provided by the present invention;

[0028] Figure 3 A schematic diagram of a channel feature learning process provided by the present invention;

[0029] Figure 4 This is a schematic diagram of a joint channel prediction and beamforming optimization device based on a Fourier graph neural network provided by the present invention. Detailed Implementation

[0030] The present invention will be further described in detail below with reference to specific embodiments, but the implementation of the present invention is not limited thereto.

[0031] To effectively address the channel aging problem and improve beamforming accuracy and system performance, this invention proposes a joint channel prediction and beamforming optimization method based on a Fourier Graph Neural Network (FourierGNN). This method transforms the input historical channel information into a hypervariable graph, learning the spatiotemporal correlations between channels in a purely graph-based manner. Matrix multiplication is performed in Fourier space using stacked Fourier Graph Operators (FGOs). This invention effectively alleviates the channel aging problem, optimizes the beamforming matrix solution process, and maximizes system performance and rate with low complexity. This method considers the spatiotemporal correlations of channels in a multi-user MISO system holistically, thereby mitigating the CSI aging problem. Furthermore, it simplifies the complex operations in directly solving the beamforming process through an optimal solution structure, resulting in a more efficient and accurate beamforming matrix.

[0032] Figure 1 This invention provides a flowchart illustrating a joint channel prediction and beamforming optimization method based on Fourier graph neural networks, as shown below. Figure 1 As shown, the method includes:

[0033] S101. Based on the signal received by the base station, estimate the historical channel state information of L consecutive timestamps. The signal received by the base station is the signal sent by the user to the base station based on the known pilot signal.

[0034] S102. Input the historical channel state information with L consecutive timestamps into the target network through a hybrid training method to obtain the downlink beamforming matrix.

[0035] The target network includes a channel prediction module, a power learning module, and a beamforming recovery module.

[0036] The channel prediction module is used to predict the channel state information at any future time based on L consecutive historical channel state information timestamps.

[0037] The power learning module is used to obtain the power characteristics of future time moments based on the predicted channel state information.

[0038] The beamforming recovery module is used to obtain the downlink beamforming matrix based on the channel state information, power characteristics, and optimal beamforming solution structure at future time moments.

[0039] The entire network consists of three key modules: a channel prediction module, a power learning module, and a beamforming recovery module. These modules work together to maximize the overall system speed and performance while meeting transmit power constraints.

[0040] This invention provides a joint channel prediction and beamforming optimization method based on Fourier graph neural networks. It estimates L consecutive historical channel state information at L timestamps based on signals received from the base station. This L consecutive historical channel state information is then input into a target network trained using a hybrid training method to obtain a downlink beamforming matrix. The target network includes a channel prediction module, a power learning module, and a beamforming recovery module. The channel prediction module predicts the channel state information at any future time based on the L consecutive historical channel state information. The power learning module obtains the power characteristics of the future time based on the predicted channel state information. The beamforming recovery module obtains the downlink beamforming matrix based on the future channel state information, the future power characteristics, and the optimal beamforming solution structure. This method considers the spatial correlation in multi-user, multi-antenna systems from a spatiotemporally dynamic perspective. Furthermore, it optimizes beamforming while addressing the channel timeout problem, reducing computational complexity, accelerating convergence, and achieving the goal of maximizing system performance and data rate.

[0041] In one possible implementation, the channel prediction module is specifically used to perform the following operations:

[0042] 1. Construct a spatiotemporal fully connected graph based on L consecutive historical channel state information with timestamps.

[0043] The corresponding expression is:

[0044]

[0045] in, Represents a fully connected graph in spacetime. Representing nodes, specifically time series. any element in, It contains L timestamps, and each timestamp corresponds to N feature variables. Let represent an N×L real matrix. This represents the adjacency matrix.

[0046] 2. A d-dimensional vector was assigned to each node in the spatiotemporally fully connected graph, and each node was mapped to the node embedding matrix according to the assignment result.

[0047] The corresponding expression is:

[0048]

[0049] in, Represents the node embedding matrix. This represents a d-dimensional vector.

[0050] 3. Channel features are learned through Fourier GNN graph neural network, based on spatiotemporal fully connected graph and mapped node embedding matrix.

[0051] 4. By using a two-layer feedforward network, the learned channel features are mapped to future times to obtain the predicted value of the channel state information at any future time.

[0052] The corresponding expression is:

[0053]

[0054] in, This represents the predicted channel value from the base station to user k at time t+n, and FFN() represents the feedforward neural network. This represents the inverse Fourier transform. This represents the output of FourierGNN, where n represents a constant.

[0055] Further optional, the graph neural network FourierGNN includes a discrete Fourier transform unit, an FGO stacked unit, and a discrete Fourier inverse transform unit.

[0056] Accordingly, channel features are learned using a Fourier GNN graph neural network based on a spatiotemporally fully connected graph and the mapped node embedding matrix, including the following steps A1 and A2:

[0057] A1. Construct the Green's kernel function using the Discrete Fourier Transform (DFT) unit, and then use the DFT to transform the Green's kernel function from the time domain to the frequency domain, obtaining the Fourier plot operator. The corresponding expression is:

[0058]

[0059] Where κ represents the Green's kernel function, [NL] represents the node index set, and R d×d Let represent a d×d real matrix. This represents the Fourier graph operator. Represents the Discrete Fourier Transform. Let represent an NL×d×d complex matrix.

[0060] A2. Using FGO stacked units, perform arithmetic operations on the Fourier plot operator in the frequency domain to obtain... The corresponding expression is:

[0061]

[0062] Where M represents the total number of FGO layers in the FGO stacking unit, σ() represents the activation function, and S i Let FGO be the i-th layer, where i represents the i-th layer. This represents the output of FourierGNN, where FourierGNN() represents FourierGNN, m represents the FGO of the m-th layer, and S 0:m b represents the product of FGO from level 0 to level m. m Indicates the bias term. Represents the Discrete Fourier Transform. Let W represent the adjacency matrix of the i-th layer. i Let represent the weight matrix of the i-th layer.

[0063] Optionally, the loss function for the channel prediction module is:

[0064]

[0065] Where I represents the total number of prediction channels in the training set. and Let represent the predicted channel and the actual channel of the i-th sample, respectively.

[0066] Specifically, firstly, this application defines a new data structure called a spatiotemporal fully connected graph, or hypervariable graph. Assuming a time series... It contains L timestamps, each corresponding to N feature variables. The hypervariable graph treats each element in X as a node. The connection strength between different nodes is represented by an adjacency matrix. These nodes... and adjacency matrix Together, they form a fully connected spatiotemporal graph, represented as:

[0067]

[0068] in, Represents a fully connected graph in spacetime. Representing nodes, specifically time series. any element in, It contains L timestamps, and each timestamp corresponds to N feature variables. Let represent an N×L real matrix. This represents the adjacency matrix.

[0069] based on The channel prediction problem can be reformulated as:

[0070]

[0071] in, This represents the parameters of the graph neural network.

[0072] To capture both local and global information about nodes and enhance the model's representational capabilities, a d-dimensional vector is assigned to each node. Mapping to node embedding matrix middle, It can be represented as:

[0073]

[0074] in, This represents the embedding matrix.

[0075] Traditionally, graph convolution operations should be used to handle graph structures. However, as the number of variables N increases and the time window L expands, this approach may encounter computational bottlenecks.

[0076] like Figure 3 As shown, to address this problem, this invention introduces a novel graph neural network, FourierGNN. This network introduces a weight matrix... To obtain a custom Green's kernel function. The Green's kernel function effectively captures the complex interactions between node features, and it is expressed as:

[0077]

[0078] Where κ represents the Green's kernel function, and [NL] represents the node index set. Let represent a d×d real matrix. This function combines node feature weights and maps the connectivity between two nodes onto a d×d matrix, thus providing richer information about node interactions. For each pair of nodes [i,j], define:

[0079]

[0080] κ[i,j]=κ[ij],

[0081] in, It is the Hadama product.

[0082] Subsequently, by employing the Discrete Fourier Transform (DFT) to transform the function κ from the time domain to the frequency domain, the convolution operation can be simplified to a dot product operation. This transformation significantly simplifies computational complexity, especially when dealing with large-scale graphs and high-dimensional features. The transformed result is called the Fourier Graph Operator (FGO), denoted as:

[0083]

[0084] in, This represents the Fourier graph operator. Represents the Discrete Fourier Transform. Let represent an NL×d×d complex matrix.

[0085] Based on the properties of DFT and the definition of convolution, the following derivation can be obtained:

[0086]

[0087] Where * represents convolution operation.

[0088] and

[0089]

[0090] Therefore, the convolution formula can be obtained:

[0091]

[0092] In the frequency domain and The multiplication operation corresponds to the graph convolution operation in the time domain, but the former has a much lower complexity than the latter.

[0093] The FourierGNN proposed in this application is implemented by stacking multiple FGOs in the frequency domain, and its mathematical expression is as follows:

[0094]

[0095] Where M represents the total number of FGO layers in the FGO stacking unit, σ() represents the activation function, and S i Let FGO be the i-th layer, where i represents the i-th layer. This represents the output of FourierGNN, where FourierGNN() represents FourierGNN, m represents the FGO of the m-th layer, and S 0:m b represents the product of FGO from level 0 to level m. m Indicates the bias term. This represents the Discrete Fourier Transform.

[0096] Finally, the Discrete Fourier Inverse Transform (IDFT) is used to... The frequency domain is transformed into the time domain, and then mapped to future times using a two-layer feed-forward network (FFN), ultimately yielding the following equation:

[0097]

[0098] in, This represents the predicted channel value from the base station to user k at time t+n, and FFN() represents the feedforward neural network. This represents the inverse Fourier transform. This represents the output of FourierGNN, where n represents a constant.

[0099] The loss function of the prediction network is:

[0100]

[0101] Where I represents the total number of prediction channels in the training set. and Let represent the predicted channel and the true channel of the i-th sample, respectively, and ‖·‖ represent the Euclidean norm.

[0102] By using FourierGNN to predict channels from a spatiotemporally dynamic perspective, the complex characteristics between channels are learned, mitigating the channel obsolescence problem and improving beamforming accuracy.

[0103] Optionally, the optimal beamforming solution structure is expressed as:

[0104]

[0105] Where I represents the identity matrix, and p = [p1, ..., p2] K ] T Let λ represent the downlink power allocation vector, where λ = [λ1, ..., λ2]. K ] T The uplink power is represented by [p,λ], which determines the beamforming direction. [p,λ] satisfies... The superscript H indicates conjugate transpose, and the superscript T indicates transpose.

[0106] It is easy to see that the above optimal solution structure transforms the maximization and rate problem into the problem of finding optimal power features [p, λ], which can be learned through a neural network. Their true values ​​can be obtained using the WMMSE algorithm. This approach can reduce the output dimension of the neural network from 2N. t K is reduced to 2K, thus accelerating training and testing. The beamforming vector can then be recovered using the aforementioned optimal solution structure. The joint channel prediction and beamforming optimization method based on Fourier graph neural network proposed in this invention can directly predict the beamforming matrix based on historical channel information and effectively handle the CSI outdated problem.

[0107] Based on deep learning methods, the proposed joint channel prediction and beamforming optimization method based on Fourier graph neural network is a joint optimization strategy. It transforms the complex beamforming optimization problem into a parameter learning problem through the optimal solution structure, and then directly predicts the optimal beamforming matrix, which greatly reduces the computational complexity.

[0108] Furthermore, in one possible implementation, the loss function obtained when training the target network using a hybrid training method is expressed as:

[0109] L = L H +L P +βL R ,

[0110] Among them, L R L represents the loss function obtained through unsupervised learning. H L represents the loss function of the channel prediction module. P L represents the loss function of the power learning module. H and L P Both loss functions are obtained through supervised learning, where β is a positive coefficient.

[0111]

[0112] The predicted value represents the channel state information. Represents the downlink beamforming vector.

[0113] R sum () represents the system and rate, R sum The following conditions must be met:

[0114]

[0115] P represents the maximum transmit power of the base station, w k (t+n) represents the downlink beamforming vector of user k at time t+n, R sum (t+n) represents the system and speed, k represents the user index, and K represents the total number of users. Let |||| represent the signal-to-interference-plus-noise ratio (SIR) at user k at time t+n, where t+n represents time. 2 dl represents the square of the Euclidean norm, and dl represents the downlink.

[0116] By employing a hybrid training method, the summation rate can be maximized through the summation rate expression. Adding a loss function helps optimize the training results of the entire neural network and improve system performance.

[0117] Optionally, the power learning module is a multilayer perceptron.

[0118] Specifically, the input first obtains the CSI at time t+n through the channel prediction module. Then, The data is fed into the power learning module, where a multilayer perceptron (MLP) can be used to obtain the power features corresponding to time t+n. The power learning module can use mean square error (MSE) as the loss function.

[0119] Correspondingly, the optional loss function for the power learning module is:

[0120]

[0121] in, and Let these represent the actual downlink and uplink power values ​​of the k-th user in the i-th sample, respectively. and Let represent the predicted downlink and uplink power values ​​for the k-th user in the i-th sample, respectively.

[0122] Figure 4 This is a schematic diagram of the structure of a joint channel prediction and beamforming optimization device based on a Fourier graph neural network provided by the present invention, as shown below. Figure 4 As shown, the device includes:

[0123] The acquisition module 41 is used to estimate L consecutive historical channel state information based on the signal received by the base station. The signal received by the base station is the signal sent by the user to the base station based on the known pilot signal.

[0124] Processing module 42 is used to input the historical channel state information with L consecutive timestamps into the target network trained by the hybrid training method to obtain the downlink beamforming matrix.

[0125] The target network includes a channel prediction module, a power learning module, and a beamforming recovery module.

[0126] The channel prediction module is used to predict the channel state information at any future time based on L consecutive historical channel state information timestamps.

[0127] The power learning module is used to obtain the power characteristics of future time moments based on the predicted channel state information.

[0128] The beamforming recovery module is used to obtain the downlink beamforming matrix based on the channel state information, power characteristics, and optimal beamforming solution structure at future time moments.

[0129] The present invention also provides an electronic device structure, including a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus.

[0130] Memory, used to store computer programs;

[0131] When a processor executes a program stored in memory, it implements the steps provided in the above method embodiments.

[0132] The communication interface is used for communication between the aforementioned electronic devices and other devices.

[0133] The method provided in this invention can be applied to electronic devices. Specifically, the electronic device can be a desktop computer, a portable computer, a smart mobile terminal, a server, etc. No limitation is made herein; any electronic device that can implement this invention falls within the protection scope of this invention.

[0134] For the device / electronic device embodiments, since they are basically similar to the method embodiments, the description is relatively simple. For specific details and beneficial effects, please refer to the description of the method embodiments.

[0135] The terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0136] The above description, in conjunction with specific preferred embodiments, provides a further detailed explanation of the present invention. It should not be construed that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, various simple deductions or substitutions can be made without departing from the concept of the present invention, and all such modifications and substitutions should be considered within the scope of protection of the present invention.

Claims

1. A joint channel prediction and beamforming optimization method based on Fourier graph neural networks, characterized in that, include: Based on the signal received by the base station, the following can be estimated: The base station receives continuous historical channel state information with timestamps, and the signal received by the base station is a signal sent by the user to the base station based on a known pilot signal. The The continuous historical channel state information with timestamps is input into the target network and trained using a hybrid training method to obtain the downlink beamforming matrix; The target network includes a channel prediction module, a power learning module, and a beamforming recovery module. The channel prediction module is used to predict based on the By using continuous historical channel state information with timestamps, the channel state information at any future moment can be predicted. The power learning module is used to obtain the power characteristics of the future time based on the predicted channel state information. The beamforming recovery module is used to obtain the downlink beamforming matrix based on the channel state information at the future time, the power characteristics at the future time, and the optimal beamforming solution structure.

2. The method according to claim 1, characterized in that, The loss function obtained by training the target network using a hybrid training method is expressed as: ， in, This represents the loss function obtained through unsupervised learning. This represents the loss function of the channel prediction module. This represents the loss function of the power learning module. and Both loss functions were obtained through supervised learning. It is a positive coefficient. ， The predicted value representing channel state information. Represents the downlink beamforming vector. Indicates system and rate, The following conditions must be met: ， This indicates the maximum transmit power of the base station. express Time users downlink beamforming vector, Indicates system and rate, Indicates the user's index. This represents the total number of users. express Time users Signal-to-noise ratio at the location Indicates time, To represent the square of the Euclidean norm, Indicates the downlink.

3. The method according to claim 2, characterized in that, The channel prediction module is specifically used for: According to the above Using continuous historical channel state information with timestamps, a spatiotemporally fully connected graph is constructed, and the corresponding expression is: ， in, Represents a fully connected graph in spacetime. Representing nodes, specifically time series. any element in, Includes Each timestamp corresponds to a timestamp. One characteristic variable, Represent a A real matrix, Represents the adjacency matrix; A node was assigned to each node corresponding to the spatiotemporally fully connected graph. The dimensional vector is used to map each node to a node embedding matrix based on the allocation result. The corresponding expression is: ， in, Represents the embedding matrix. Represents the node embedding matrix. express dimensional vector; Channel features are learned using a Fourier GNN graph neural network based on the spatiotemporally fully connected graph and the mapped node embedding matrix. By using a two-layer feedforward network, the learned channel features are mapped to future times to obtain the predicted value of the channel state information at any future time. The corresponding expression is: ， in, express Time base station to user Channel prediction value, This represents a feedforward neural network. This represents the inverse Fourier transform. This represents the output of FourierGNN. Represents a constant.

4. The method according to claim 3, characterized in that, The graph neural network FourierGNN includes a discrete Fourier transform unit, an FGO stacked unit, and a discrete Fourier inverse transform unit. Accordingly, the learning of channel features using a Fourier GNN graph neural network, based on the spatiotemporally fully connected graph and the mapped node embedding matrix, includes: The Green's kernel function is constructed using the Discrete Fourier Transform (DFT) unit, and then transformed from the time domain to the frequency domain using the DFT to obtain the Fourier plot operator, the corresponding expression of which is: ， ， in, Represents the Green kernel function. Represents the set of indexes for nodes. Represent a A real matrix, This represents the Fourier graph operator. Represents the Discrete Fourier Transform. Represent a Complex matrices; Through the FGO stacking unit, arithmetic operations are performed on the Fourier plot operator in the frequency domain to obtain... The corresponding expression is: ， ， in, This indicates the total number of FGO stacks in the FGO stack unit. This represents the activation function. Indicates the first FGO of the layer, Indicates the first layer, This represents the output of FourierGNN. This represents FourierGNN. Indicates the first FGO of the layer, This represents the product of FGO from level 0 to level m. Indicates the bias term. Represents the Discrete Fourier Transform. , Indicates the first The adjacency matrix of the layer, Indicates the first Layer weight matrix.

5. The method according to claim 4, characterized in that, The loss function of the channel prediction module is: ， in, This represents the total number of prediction channels in the training set. and They represent the first The predicted channel and the real channel for each sample.

6. The method according to claim 5, characterized in that, The loss function of the power learning module is: ， in, and They represent the first In the nth sample The actual values ​​of downlink and uplink power for each user. and They represent the first In the nth sample Predicted values ​​for downlink and uplink power for each user. This represents the total number of prediction channels in the training set.

7. The method according to claim 5, characterized in that, The power learning module is a multilayer perceptron.

8. A joint channel prediction and beamforming optimization device based on Fourier graph neural networks, characterized in that, include: The acquisition module is used to estimate the signal received by the base station. The base station receives continuous historical channel state information with timestamps, and the signal received by the base station is a signal sent by the user to the base station based on a known pilot signal. Processing module, used to process the The continuous historical channel state information with timestamps is input into the target network and trained using a hybrid training method to obtain the downlink beamforming matrix; The target network includes a channel prediction module, a power learning module, and a beamforming recovery module. The channel prediction module is used to predict based on the By using continuous historical channel state information with timestamps, the channel state information at any future moment can be predicted. The power learning module is used to obtain the power characteristics of the future time based on the predicted channel state information. The beamforming recovery module is used to obtain the downlink beamforming matrix based on the channel state information at the future time, the power characteristics at the future time, and the optimal beamforming solution structure.

9. An electronic device, characterized in that, It includes a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus; Memory, used to store computer programs; A processor, when executing a program stored in memory, implements the method described in any one of claims 1-7.

Images