Joint channel prediction and beam forming optimization method based on Fourier graph neural network

Abstract

The invention provides a joint channel prediction and beam forming optimization method based on a Fourier graph neural network, and the method comprises the steps: carrying out the estimation of L pieces of historical channel state information with continuous timestamps according to a signal received by a base station, inputting the information into a target network obtained through the training of a mixed training method, obtaining a downlink beam forming matrix; the target network comprises a channel prediction module, a power learning module and a beam forming recovery module; the channel prediction module is used for predicting to obtain channel state information at any future moment according to L historical channel state information with continuous timestamps; the power learning module is used for obtaining power characteristics at the future moment according to the predicted channel state information at the future moment; and the beam forming recovery module is used for obtaining a downlink beam forming matrix according to the channel state information at the future moment, the power characteristics at the future moment and the beam forming optimal solution structure, optimizing a channel outdated problem and beam forming, and realizing system performance and rate maximization.

Description

Technical Field

[0001] The present invention relates to the field of communication technology, and in particular to a joint channel prediction and beamforming optimization method based on Fourier graph neural network. Background Art

[0002] Beamforming technology has the advantages of signal focusing and interference suppression, and has shown excellent performance in the field of wireless communications. This performance depends largely on the accuracy and real-time performance of channel state information (CSI). The more accurate the channel, the more effective the beamforming. However, due to the dynamics of the wireless channel and the feedback delay of CSI, the CSI obtained at the base station is seriously outdated, which greatly reduces the beamforming effect and thus leads to a serious degradation of system performance. In addition, traditional algorithms that directly solve the beamforming matrix through channel state CSI (such as the weighted minimum mean squared error (WMMSE) algorithm) usually require multiple iterations to find the local optimal solution, which may lead to problems such as high computational complexity and long convergence time, especially in high-dimensional problems, where a large number of iterations are required to converge to the optimal solution. To solve these problems, joint channel prediction and beamforming design have become the key to optimizing communication performance.

[0003] Traditional channel prediction methods usually use models such as Recurrent Neural Networks (RNN), Long Short-Term Memory (LSTM) networks, and Gated Recurrent Units (GRU) to make predictions, and are committed to studying the temporal correlation between channels. However, for multi-user multiple-input single-output (MISO) systems, the channel response of each user usually also has significant spatial correlation, and the simple temporal correlation may not be sufficient to predict accurate channel information. There are also some emerging methods that use graph neural networks to deal with the channel staleness problem, but these methods either rely heavily on pre-defined graph structures to specify spatial correlations, but in fact cannot capture spatial correlations that change over time, or stack graph networks (such as Graph Convolutional Network (GCN) and Graph Attention Network (GAT)) and temporal networks (such as LSTM and GRU) to capture spatial dynamics and temporal dependencies respectively. This separate modeling of spatiotemporal features contradicts the unity of spatiotemporal correlations in the real world, which seriously affects the prediction performance. Besides this, to date, relatively few studies have considered both channel prediction and beamforming design simultaneously.

[0004] The channel prediction method based on Spectral-Temporal Graph Neural Network (StemGNN) uses spectrum sequence units and graph convolution to capture the spatiotemporal correlation of channels in the frequency domain respectively. However, this invention learns temporal characteristics and spatial characteristics separately. This idea contradicts the theory of spatiotemporal dynamic unification in the real world, which seriously affects the prediction performance. In addition, this technology does not consider the optimization problem of beamforming technology. Another method in the prior art is to solve the beamforming problem based on the LSTM joint learning framework and consider the outdated CSI, but due to the limitations of LSTM, its performance needs to be improved. Summary of the invention

[0005] In order to solve the above problems existing in the prior art, the present invention provides an ordered statistics decoding method, device and apparatus based on a belief propagation list, which specifically include:

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

[0007] According to the signal received by the base station, historical channel state information with L consecutive timestamps is estimated, 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;

[0008] Inputting L historical channel state information with continuous timestamps into the target network trained by the hybrid training method to obtain a downlink beamforming matrix;

[0009] Among them, the target network includes a channel prediction module, a power learning module and a beamforming recovery module;

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

[0011] A power learning module is used to obtain the channel state information at a future time according to the prediction and obtain the power characteristics at a future time;

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

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

[0014] An acquisition module, used to estimate and obtain historical channel state information with L consecutive timestamps according to a signal received by the base station, where the signal received by the base station is a signal sent by a user to the base station based on a known pilot signal;

[0015] A processing module, used for inputting historical channel state information with L consecutive timestamps into a target network trained by a hybrid training method to obtain a downlink beamforming matrix;

[0016] Among them, the target network includes a channel prediction module, a power learning module and a beamforming recovery module;

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

[0018] A power learning module is used to obtain the channel state information at a future time according to the prediction and obtain the power characteristics at a future time;

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

[0020] In a third aspect, the present invention further provides an electronic device, comprising a processor, a communication interface, a memory and a communication bus, wherein the processor, the communication interface and the memory communicate with each other via the communication bus;

[0021] Memory, used to store computer programs;

[0022] The processor is used to implement any method provided in the first aspect when executing a program stored in the memory.

[0023] Beneficial effects of the present invention:

[0024] The joint channel prediction and beamforming optimization method based on Fourier graph neural network provided by the present invention estimates and obtains historical channel state information with L consecutive timestamps according to the signal received by the base station; inputs the historical channel state information with L consecutive timestamps into the target network trained by the hybrid training method to obtain a downlink beamforming matrix; wherein 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 the channel state information at any future time according to the historical channel state information with L consecutive timestamps; the power learning module is used to obtain the power characteristics at the future time according to the predicted channel state information at the future time; the beamforming recovery module is used to obtain the downlink beamforming matrix according to the channel state information at the future time, the power characteristics at the future time and the optimal solution structure of the beamforming, and the spatial correlation existing in the multi-user multi-antenna system is considered from the perspective of the dynamic unification of time and space, and while optimizing the channel obsolescence problem, the beamforming is optimized, the calculation complexity is reduced, the convergence speed is accelerated, and the goal of maximizing the system and rate is achieved.

[0025] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 A schematic diagram of a process flow of a joint channel prediction and beamforming optimization method based on Fourier graph neural network provided by the present 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 A structural schematic diagram of a joint channel prediction and beamforming optimization device based on Fourier graph neural network provided by the present invention. DETAILED DESCRIPTION

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

[0031] In order to effectively solve the channel obsolescence problem and improve the accuracy and system performance of beamforming, the present invention proposes a joint channel prediction and beamforming optimization method based on Fourier Graph Neural Network (FourierGNN), which converts the input historical channel information into the form of a hypervariable graph, learns the spatiotemporal correlation between channels in a pure graph manner, and performs matrix multiplication operations in Fourier space by stacking Fourier Graph Operators (FGOs). The present invention can well alleviate the channel obsolescence problem, optimize the solution process of the beamforming matrix, and maximize the system and rate under low complexity. The method considers the spatiotemporal correlation of channels in a multi-user MISO system as a whole, thereby alleviating the CSI obsolescence problem, and simplifies the various complex operations in the direct solution of the beamforming process through the optimal solution structure, thereby obtaining a more efficient and accurate beamforming matrix.

[0032] Figure 1 A schematic diagram of a process flow of a joint channel prediction and beamforming optimization method based on Fourier graph neural network provided by the present invention is shown in FIG. Figure 1 As shown, the method includes:

[0033] S101. Estimate and obtain historical channel state information with L consecutive time stamps based on a signal received by a base station. The signal received by the base station is a signal sent by a user to the base station based on a known pilot signal.

[0034] S102: inputting L historical channel state information with consecutive time stamps into a target network trained by a hybrid training method to obtain a 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 according to the historical channel state information of L consecutive time stamps.

[0037] The power learning module is used to obtain the channel state information at a future time according to the prediction and obtain the power characteristics at the future time.

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

[0039] The entire network consists of three key modules, namely the channel prediction module, the power learning module and the beamforming recovery module, which work together to maximize the sum rate and performance of the entire system while meeting the transmission power constraints.

[0040] The joint channel prediction and beamforming optimization method based on Fourier graph neural network provided by the present invention estimates and obtains historical channel state information with L consecutive timestamps according to the signal received by the base station; inputs the historical channel state information with L consecutive timestamps into the target network trained by the hybrid training method to obtain a downlink beamforming matrix; wherein 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 the channel state information at any future time according to the historical channel state information with L consecutive timestamps; the power learning module is used to obtain the power characteristics at the future time according to the predicted channel state information at the future time; the beamforming recovery module is used to obtain the downlink beamforming matrix according to the channel state information at the future time, the power characteristics at the future time and the optimal solution structure of the beamforming, and the spatial correlation existing in the multi-user multi-antenna system is considered from the perspective of the dynamic unification of time and space, and while optimizing the channel obsolescence problem, the beamforming is optimized, the calculation complexity is reduced, the convergence speed is accelerated, and the goal of maximizing the system and rate is achieved.

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

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

[0043] The corresponding expression is:

[0044]

[0045] in, represents a spatiotemporal fully connected graph, Represents a node, specifically a time series Any element in contains L timestamps, each of which corresponds to N feature variables. represents an N×L real matrix, Represents the adjacency matrix.

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

[0047] The corresponding expression is:

[0048]

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

[0050] 3. Through the graph neural network FourierGNN, channel features are learned based on the spatiotemporal fully connected graph and the mapped node embedding matrix.

[0051] 4. Through a double-layer feedforward network, the learned channel features are mapped to future moments to obtain the predicted value of the channel state information at any future moment.

[0052] The corresponding expression is:

[0053]

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

[0055] Further optionally, a graph neural network FourierGNN includes a discrete Fourier transform unit, a FGO stacking unit and a discrete Fourier inverse transform unit.

[0056] Accordingly, the graph neural network FourierGNN is used to learn channel features based on the spatiotemporal fully connected graph and the mapped node embedding matrix, including the following steps A1 and A2:

[0057] A1. Through the discrete Fourier transform unit, the Green kernel function is constructed, and the discrete Fourier transform is used to convert the Green kernel function from the time domain to the frequency domain to obtain the Fourier graph operator. The corresponding expression is:

[0058]

[0059] Among them, κ represents the Green kernel function, [NL] represents the index set of the node, and R d×d represents a d×d real matrix, represents the Fourier graph operator, represents the discrete Fourier transform, Represents a NL×d×d complex matrix.

[0060] A2. Through the FGO stacking unit, arithmetic operations are performed on the Fourier graph 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 represents the FGO of the i-th layer, i represents the i-th layer, represents the output of FourierGNN, FourierGNN() represents FourierGNN, m represents the FGO of the mth layer, S 0:m represents the product of FGO from the 0th layer to the mth layer, b m represents the bias term, represents the discrete Fourier transform, represents the adjacency matrix of the i-th layer, W i represents the weight matrix of the i-th layer.

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

[0064]

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

[0066] Specifically, first, this application defines a new data structure called a spatiotemporal fully connected graph, i.e., a hypervariable graph. Assume that the time series Contains L timestamps, each of which corresponds to N feature variables. The hypervariable graph regards each element in X as a node. The connection strength between different nodes is represented by the adjacency matrix. These nodes And the adjacency matrix Together they form a spatiotemporal fully connected graph, expressed as:

[0067]

[0068] in, represents a spatiotemporal fully connected graph, Represents a node, specifically a time series Any element in contains L timestamps, each of which corresponds to N feature variables. represents an N×L real matrix, Represents the adjacency matrix.

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

[0070]

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

[0072] In order to capture the local and global information of the node and enhance the representation ability of the model, a d-dimensional vector is assigned to each node. Mapping to node embedding matrix middle, It can be expressed as:

[0073]

[0074] in, represents the embedding matrix.

[0075] According to traditional methods, graph convolution operations should be used to process graph structures. However, this approach may face computational bottlenecks as the number of variables N increases and the time window L expands.

[0076] like Figure 3 As shown, in order to solve this problem, the present invention introduces a new graph neural network FourierGNN. The network introduces the weight matrix To obtain a custom Green's kernel function. The Green's kernel function can effectively capture the complex interactions between node features, which is expressed as:

[0077]

[0078] Among them, κ represents the Green kernel function, [NL] represents the index set of the node, Represents a d×d real matrix. This function combines the node feature weights and maps the connectivity between two nodes to a d×d matrix, thus providing richer node interaction information. For each pair of nodes [i, j], define:

[0079]

[0080] κ[i,j]=κ[ij],

[0081] in, is the Hadamard product.

[0082] Subsequently, by using 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 point multiplication operation. This transformation greatly simplifies the computational complexity, especially when dealing with large-scale graphs and high-dimensional features. The result after the transformation is called Fourier Graph Operator (FGO), which is expressed as:

[0083]

[0084] in, represents the Fourier graph operator, represents the discrete Fourier transform, Represents a NL×d×d complex matrix.

[0085] According to the properties of DFT and the definition of convolution operation, the following derivation can be obtained:

[0086]

[0087] Among them, * represents the 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, and the complexity of the former is much lower 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:

[0094]

[0095] Where M represents the total number of FGO layers in the FGO stacking unit, σ() represents the activation function, and S i represents the FGO of the i-th layer, i represents the i-th layer, represents the output of FourierGNN, FourierGNN() represents FourierGNN, m represents the FGO of the mth layer, S 0:m represents the product of FGO from the 0th layer to the mth layer, b m represents the bias term, represents discrete Fourier transform.

[0096] Finally, the Discrete Fourier Inverse Transform (IDFT) is used to transform Convert from frequency domain to time domain and map it to the future time using a two-layer feed-forward network (FFN), and finally get the following formula:

[0097]

[0098] in, represents the channel prediction value from the base station to user k at time t+n, FFN() represents the feedforward neural network, represents the inverse Fourier transform, Represents the output of FourierGNN, and 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 denote the predicted channel and true channel of the i-th sample respectively, and ‖·‖ represents the Euclidean norm.

[0102] Through FourierGNN, channel prediction is performed from a unified perspective of spatiotemporal dynamics, the complex characteristics between channels are learned, and the channel obsolescence problem is alleviated to improve the accuracy of beamforming.

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

[0104]

[0105] Where I represents the identity matrix, p = [p1,…,p K ] T represents the downlink power allocation vector, λ=[λ1,…,λ K ] T represents the uplink power, which determines the direction of beamforming, and [p,λ] satisfies The superscript H denotes conjugate transpose, and the superscript T denotes transpose.

[0106] It is not difficult to see that the above optimal solution structure transforms the maximization and rate problem into the problem of finding the optimal power characteristics [p,λ], which can be learned by the neural network. Their true values ​​can be obtained by the WMMSE algorithm. This approach can reduce the output dimension of the neural network from 2N t K is reduced to 2K, which speeds up training and testing. Then the beamforming vector can be restored through the above optimal solution structure. The joint channel prediction and beamforming optimization method based on Fourier graph neural network proposed in the present invention can directly predict the beamforming matrix according to historical channel information and effectively deal with the CSI outdated problem.

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

[0108] Furthermore, in a possible implementation, the loss function of the target network obtained by training with the hybrid training method is expressed as:

[0109] L=L H +L P +βL R ,

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

[0111]

[0112] represents the predicted value of 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 transmission 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 sum rate, k represents the index of the user, K represents the total number of users, represents the signal-to-interference-noise ratio at user k at time t+n, t+n represents the time, |||| 2 represents the square of the Euclidean norm, and dl represents downlink.

[0116] Through the hybrid training method, the sum rate can be maximized by expressing the sum rate as 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 multi-layer perceptron.

[0118] Specifically, the input first passes through the channel prediction module to obtain the CSI at the future time t+n Then, It is sent to 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 of the power learning module is:

[0120]

[0121] in, and denote the true value of the downlink and uplink power of the kth user in the ith sample, respectively, and They represent the predicted values ​​of downlink and uplink power of the kth user in the ith sample, respectively.

[0122] Figure 4 A structural schematic diagram of a joint channel prediction and beamforming optimization device based on Fourier graph neural network provided by the present invention, such as Figure 4 As shown, the device comprises:

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

[0124] The processing module 42 is used to input the historical channel state information with L consecutive time stamps into the target network trained by the hybrid training method to obtain a 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 according to the historical channel state information of L consecutive time stamps.

[0127] The power learning module is used to obtain the channel state information at a future time according to the prediction and obtain the power characteristics at the future time.

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

[0129] The present invention also provides a structure of 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 via the communication bus.

[0130] Memory, used to store computer programs;

[0131] The processor is used to implement the steps provided in the above method embodiment when executing the program stored in the memory.

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

[0133] The method provided in the embodiment of the present invention can be applied to electronic devices. Specifically, the electronic device can be: a desktop computer, a portable computer, an intelligent mobile terminal, a server, etc. This is not limited here, and any electronic device that can implement the present invention belongs to the protection scope of the present invention.

[0134] As for the device / electronic device embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the specific contents and beneficial effects and other related matters can be referred to the partial description of the method embodiment.

[0135] The terms "first" and "second" are used for descriptive purposes only and should not be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the features. In the description of the present invention, the meaning of "plurality" is two or more, unless otherwise clearly and specifically defined.

[0136] The above contents are further detailed descriptions of the present invention in combination with specific preferred embodiments, and it cannot be determined that the specific implementation of the present invention is limited to these descriptions. For ordinary technicians in the technical field to which the present invention belongs, several simple deductions or substitutions can be made without departing from the concept of the present invention, which should be regarded as falling within the protection scope of the present invention.

Claims

1. A joint channel prediction and beamforming optimization method based on Fourier graph neural network, characterized in that: include: According to a signal received by the base station, historical channel state information with L consecutive timestamps is estimated, where the signal received by the base station is a signal sent by a user to the base station based on a known pilot signal; Inputting the L historical channel state information with continuous timestamps into a target network trained by a hybrid training method to obtain a downlink beamforming matrix; Wherein, 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 the channel state information at any future time according to the L consecutive historical channel state information of timestamps; The power learning module is used to obtain the power characteristics of the future moment according to the predicted channel state information of the future moment; The beamforming recovery module is used to obtain the downlink beamforming matrix according to the channel state information at the future time, the power characteristics at the future time and the beamforming optimal solution structure.

2. The method according to claim 1, characterized in that The loss function of the target network obtained by hybrid training method is expressed as: L=L H +L P +βL R , Among them, L R represents the loss function obtained by unsupervised learning, L H represents the loss function of the channel prediction module, L P represents the loss function of the power learning module, L H and L P Both are loss functions obtained through supervised learning, β is a positive coefficient, represents the predicted value of the channel state information, represents the downlink beamforming vector, R sum () represents the system and rate, R sum The following conditions must be met: P represents the maximum transmission 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 sum rate, k represents the index of the user, K represents the total number of users, represents the signal-to-interference-noise ratio at user k at time t+n, t+n represents the time, || || 2 represents the square of the Euclidean norm, and dl represents downlink.

3. The method according to claim 2, characterized in that The channel prediction module is specifically used for: According to the L historical channel state information with continuous timestamps, a spatiotemporal fully connected graph is constructed, and the corresponding expression is: in, represents a spatiotemporal fully connected graph, Represents a node, specifically a time series Any element in contains L timestamps, each of which corresponds to N feature variables. represents an N×L real matrix, represents the adjacency matrix; A d-dimensional vector is assigned to each node corresponding to the spatiotemporal fully connected graph, and each node is mapped to a node embedding matrix according to the assignment result. The corresponding expression is: in, represents the embedding matrix, represents the node embedding matrix, represents a d-dimensional vector; Learning channel features based on the spatiotemporal fully connected graph and the mapped node embedding matrix through a graph neural network FourierGNN; Through the double-layer feedforward network, the learned channel features are mapped to the future time, and the predicted value of the channel state information at any future time is obtained. The corresponding expression is: in, represents the channel prediction value from the base station to user k at time t+n, FFN() represents the feedforward neural network, represents the inverse Fourier transform, Represents the output of FourierGNN, and n 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 stacking unit and a discrete Fourier inverse transform unit; Accordingly, the graph neural network FourierGNN is used to learn channel features based on the spatiotemporal fully connected graph and the mapped node embedding matrix, including: The Green kernel function is constructed by the discrete Fourier transform unit, and the Green kernel function is converted from the time domain to the frequency domain by discrete Fourier transform to obtain the Fourier graph operator, and the corresponding expression is: Among them, κ represents the Green kernel function, [NL] represents the index set of the node, represents a d×d real matrix, represents the Fourier graph operator, represents the discrete Fourier transform, represents a NL×d×d complex matrix; Through the FGO stacking unit, the Fourier graph operator is subjected to arithmetic operation in the frequency domain to obtain The corresponding expression is: Where M represents the total number of FGO layers in the FGO stacking unit, σ() represents the activation function, and S i represents the FGO of the i-th layer, i represents the i-th layer, represents the output of FourierGNN, FourierGNN() represents FourierGNN, m represents the FGO of the mth layer, S 0:m represents the product of FGO from the 0th layer to the mth layer, b m represents the bias term, represents the discrete Fourier transform, represents the adjacency matrix of the i-th layer, W i represents the weight matrix of the i-th layer.

5. The method according to claim 4, characterized in that The loss function of the channel prediction module is: Where I represents the total number of prediction channels in the training set, and They represent the predicted channel and true channel of the i-th sample respectively.

6. The method according to claim 5, characterized in that The loss function of the power learning module is: in, and denote the true value of the downlink and uplink power of the kth user in the ith sample, respectively, and denote the predicted values ​​of downlink and uplink power of the kth user in the ith sample, respectively, and I denotes the total number of predicted channels in the training set.

7. The method according to claim 5, characterized in that The optimal beamforming solution structure is expressed as: Where I represents the identity matrix, p = [p1,…,p K ] T represents the downlink power allocation vector, λ=[λ1,…,λ K ] T represents the uplink power, which determines the direction of beamforming, and [p,λ] satisfies The superscript H denotes conjugate transpose, and the superscript T denotes transpose.

8. The method according to claim 7, characterized in that The power learning module is a multi-layer perceptron.

9. A joint channel prediction and beamforming optimization device based on Fourier graph neural network, characterized in that: include: An acquisition module, configured to estimate and obtain historical channel state information with L consecutive timestamps according to a signal received by a base station, wherein the signal received by the base station is a signal sent by a user to the base station based on a known pilot signal; A processing module, used for inputting the L historical channel state information with continuous timestamps into a target network trained by a hybrid training method to obtain a downlink beamforming matrix; Wherein, 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 the channel state information at any future time according to the L consecutive historical channel state information of timestamps; The power learning module is used to obtain the power characteristics of the future moment according to the predicted channel state information of the future moment; The beamforming recovery module is used to obtain the downlink beamforming matrix according to the channel state information at the future time, the power characteristics at the future time and the beamforming optimal solution structure.

10. 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, for implementing any of the methods described in claims 1-8 when executing a program stored in a memory.

Images