A Power Allocation Method for Downlink in a Deep Learning-Based Cellular-Free System

Abstract

This invention discloses a deep learning-based power allocation method for downlink in a cellular-free system, comprising: S1, S2 constructing a fully connected neural network structure model, transforming the problem between the power allocation variable set and the solution of traditional algorithms into a problem between the weights W and bias term b of the trained neural network, with the aim of minimizing the loss function; S3. This invention uses M-MMSE precoding to allocate power in the downlink of a large-scale MIMO network. First, a feedforward neural network with fully connected layers, consisting of an input layer, hidden layers, and an output layer, is employed to generate an estimate of the optimal power allocation vector. The neural network learns the estimates from all users to satisfy power constraints and improve estimation accuracy, significantly increasing convergence speed and reducing computational complexity.

Description

Technical Field

[0001] This invention relates to the field of power allocation methods, specifically a power allocation method for the downlink of a cellular-free system using deep learning. Background Technology

[0002] Traditional Monte Carlo-based algorithms involve a large number of high-dimensional matrix multiplications, requiring polynomial or quasi-polynomial complexity to solve and needing continuous iterative updates. However, due to the large number of antennas and users, the classic second-order algorithm is difficult to handle resource allocation due to its high computational complexity. To address this, we propose a deep learning-based power allocation method for downlink in cellular-free systems. Summary of the Invention

[0003] Therefore, the purpose of this invention is to provide a deep learning-based power allocation method for downlink in cellular-free systems. This method uses M-MMSE precoding to allocate power in the downlink of a large-scale MIMO network. First, a feedforward neural network with fully connected layers, consisting of an input layer, a hidden layer, and an output layer, is used to generate an estimate of the optimal power allocation vector. The neural network then learns the estimates from all users to satisfy power constraints and improve estimation accuracy, thus greatly improving convergence speed and reducing computational load.

[0004] To address the aforementioned technical problems, according to one aspect of the present invention, the present invention provides the following technical solution:

[0005] A power allocation method for the downlink of a deep learning-based non-cellular system, comprising:

[0006] S1: Consider the downlink of a large-scale multiple-input multiple-output network with L cells. Each cell includes a base station with M antennas and K users. A feedforward neural network with fully connected layers is used, consisting of a 2KL-dimensional input layer, N hidden layers, and a K+1-dimensional output layer, to fit the solution of all power allocation variables to the solution of a traditional algorithm. The relationship is established, and the training process is used to iterate the process step by step.

[0007] S2: Construct a fully connected neural network structure model, transforming the problem between the power allocation variable set and the solution of the traditional algorithm into a problem between the weights W and the bias term b of the trained neural network. The aim is to optimize the loss function. minimize;

[0008] S3: The dataset used by the neural network is generated by the DeepMIMO dataset. The optimal power allocation is obtained by calculation using traditional optimization methods. 80% of the samples are used for training, 10% for validation, and the remaining 10% of the samples form the test set, independent of the training set.

[0009] As a preferred embodiment of the power allocation method for the downlink of a deep learning-based non-cellular system described in this invention, the specific operation method in step S3 is as follows:

[0010] a. First, read the data of the power allocation variable set and the corresponding solution of the traditional algorithm, and then perform dimensional transformation;

[0011] b. Initialize the parameters required by the neural network;

[0012] c. Define the network structure and the number of neurons and activation function parameters in each layer. At the model level, design a 6-layer fully connected network model.

[0013] d. In the model merging stage, the Adam optimizer was used with an initial learning rate of 0.01 set in the function. The relative mean squared error was selected as the training metric, and an early stopping method was added with the patience parameter set to 50, meaning that training would be terminated if the loss function did not decrease significantly within 50 epochs.

[0014] e. During the training phase, fit the data according to the methods and parameters described above. Train using different learning rates and training-to-validation ratios, and examine the parameters to adjust hyperparameters such as learning rate, batch size, and number of training epochs. Use a small test set for prediction, and gradually adjust the parameters to obtain training results that meet the requirements.

[0015] As a preferred embodiment of the power allocation method for the downlink of a deep learning-based non-cellular system described in this invention, in the 6-layer fully connected network model, the ELU activation function is selected between the first five layers, with the following expression: x < 0: ELU(x) = e x -1, x≥0: ELU(x)=x.

[0016] Compared with the prior art, the beneficial effects of this invention are: using M-MMSE precoding to allocate power in the downlink of a large-scale MIMO network, firstly a feedforward neural network with fully connected layers, consisting of an input layer, a hidden layer and an output layer, is used to generate an estimate of the optimal power allocation vector, so that the neural network learns the estimates of all users in order to meet the power constraints and improve the estimation accuracy, which greatly improves the convergence speed and reduces the amount of computation. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the present invention will be described in detail below with reference to the accompanying drawings and detailed embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Wherein:

[0018] Figure 1 This invention provides a deep fully connected neural network structure model for a power allocation method for the downlink of a cellular-free system based on deep learning.

[0019] Figure 2 This is a transmission model diagram of a non-cellular massive MIMO system, which is a power allocation method for the downlink of a deep learning-based non-cellular system according to the present invention. Detailed Implementation

[0020] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0021] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0022] Secondly, the present invention is described in detail with reference to the schematic diagrams. When describing the embodiments of the present invention, for ease of explanation, the cross-sectional views illustrating the device structure will be partially enlarged, not at half scale. Furthermore, the schematic diagrams are merely examples and should not limit the scope of protection of the present invention. In addition, actual fabrication should include the three-dimensional spatial dimensions of length, width, and depth.

[0023] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.

[0024] This invention provides a deep learning-based power allocation method for downlink in cellular-free systems. It uses M-MMSE precoding to allocate power in the downlink of a large-scale MIMO network. First, a feedforward neural network with fully connected layers, consisting of an input layer, a hidden layer, and an output layer, is used to generate an estimate of the optimal power allocation vector. The neural network learns the estimates of all users to meet power constraints and improve estimation accuracy, which greatly improves the convergence speed and reduces the computational load.

[0025] Figure 1-2 The diagram shown is an overall structural schematic of an embodiment of a deep learning-based downlink power allocation method for cellular-free systems according to the present invention. Please refer to [link / reference]. Figure 1-2 .

[0026] 1. Neural Network Model

[0027] S1: First, consider the downlink of a large-scale multiple-input multiple-output (MLMI) network with L cells. Each cell includes a base station with M antennas and K users. A feedforward neural network with fully connected layers is used, consisting of a 2KL-dimensional input layer, N hidden layers, and a K+1-dimensional output layer, to fit the solution of all power allocation variables to the solution of a traditional algorithm. The relationship is established, and the training process is used to iterate the process step by step.

[0028] S2: Construct a fully connected neural network structure model, transforming the problem between the power allocation variable set and the solution of the traditional algorithm into a problem between the weights W and the bias term b of the trained neural network. The aim is to optimize the loss function. minimize;

[0029] S3: The dataset used by the neural network was generated from the DeepMIMO dataset. The optimal power allocation was obtained using traditional optimization methods. 80% of the samples were used for training, 10% for validation, and the remaining 10% formed the test set, independent of the training set. The specific training steps are as follows:

[0030] a. First, read the data of the power allocation variable set and the corresponding solution of the traditional algorithm, and then perform dimensional transformation;

[0031] b. Initialize the parameters required for the neural network, as detailed in the table below:

[0032]

[0033]

[0034] c. Define the network structure hierarchically, including the number of neurons and activation function parameters for each layer. At the model level, design a 6-layer fully connected network model. The specific network structure is as follows:

[0035]

[0036] d. In the model merging stage, the Adam optimizer was used with an initial learning rate of 0.01 set in the function. The relative mean squared error was selected as the training metric, and an early stopping method was added with the patience parameter set to 50, meaning that training would be terminated if the loss function did not decrease significantly within 50 epochs.

[0037] e. During the training phase, fit the data according to the methods and parameters described above. Train using different learning rates and training-to-validation ratios, and examine the parameters to adjust hyperparameters such as learning rate, batch size, and number of training epochs. Use a small test set for prediction, and gradually adjust the parameters to obtain training results that meet the requirements.

[0038] 2. System Model

[0039] Cellular-free massive MIMO breaks through the traditional design of cell-centric radio wave coverage, signal processing and resource management, and introduces an advanced user-centric network architecture design. This makes the distance between users and some APs closer, which can reduce the impact of large-scale fading, bring macro diversity gain, and provide uniform, stable and reliable user service quality in the whole area. Cellular-free network architecture can avoid frequent cell handover, reduce control signaling overhead, and save valuable radio resources.

[0040] APs are low-cost, flexible in deployment, and because each AP has a limited number of antennas, their size can be made very small, allowing for flexible deployment in space-constrained scenarios such as streetlights, manhole covers, and shelves.

[0041] The power consumption model of a cellular-free massive MIMO system consists of two parts: AP power consumption and backhaul power consumption. The power consumption of the nth AP can be modeled as follows: It can be rewritten as Where V is the set of all power allocation variables, K is the total number of users, G is the total number of multicast groups, En∈RN×N has zero elements except for [En]n, n=1, and ξn represents the efficiency of the power amplifier at the nth access point. This represents the circuit power consumption of the nth access point;

[0042] The power consumption for backhaul, used to transmit data between the access point and the CPU, is proportional to the total achievable data rate, as given by formula P. BH,n (V)=p bh,n R sum (V)+p 0,n Equation 2 gives, where, where, For the sum and rate of all services, p 0,n It is the constant power consumption for each return trip, the power consumption independent of the flow rate, while p bh,n This is power consumption related to flow rate;

[0043] Combining Equation 2, we can obtain the total power consumption of the system, which is:

[0044] By dynamically allocating the transmit power of the AP, the energy efficiency of the cellular-free massive MIMO system can be maximized. According to the definition of energy efficiency and equations 1a-1c and 3, the problem can be expressed as: st Among them, and Let be the achievable rates for multicast and unicast services, respectively; rm,g and ru,k are the minimum data rate requirements for multicast service in group g and unicast service in group k, respectively; and pn,max represents the allowed total transmit power budget for access point n.

[0045] The objective function in 4a can be rewritten as follows: Without loss of generality, maximize Equivalent to the first term of the denominator of the minimization expression (5 terms), therefore, equation 4 is equivalent to... And st(4b),(4c),(4d)-Equation 6b;

[0046] Therefore, it can be concluded that due to the non-smoothness and non-convexity of the objective function and constraints, the optimization problem given by Equation 6 is neither smooth nor convex.

[0047] Combination Figures 1-2 This embodiment presents a deep learning-based power allocation method for the downlink of a cellular-free system. In practical application, it first considers the downlink of a large-scale multiple-input multiple-output (MLMI) network with L cells. Each cell includes a base station with M antennas and K users. A feedforward neural network with fully connected layers is employed, consisting of a 2KL-dimensional input layer, N hidden layers, and a K+1-dimensional output layer, used to fit the solution of all power allocation variables to the solution of traditional algorithms. The relationship between the power allocation variable set and the solution of the traditional algorithm is transformed into the relationship between the weights W and the bias term b of the trained neural network. The aim is to improve the loss function. To minimize the impact of the dataset used in the neural network, which was generated from the DeepMIMO dataset, the optimal power allocation was calculated using traditional optimization methods. 80% of the samples were used for training, 10% for validation, and the remaining 10% formed a test set, independent of the training set. The specific steps were as follows: First, the data from the power allocation variable set was read and compared with the corresponding solutions from the traditional algorithm, and dimensionality transformation was performed. The parameters required in the neural network were initialized, and the network structure, the number of neurons in each layer, and the activation function parameters were defined hierarchically. At the model level, a 6-layer fully connected network model was designed. During the model merging phase, the Adam optimizer was used with an initial learning rate of 0.01. Relative mean squared error was chosen as the training metric, and an early stopping method was implemented with a patience parameter of 50, meaning training would terminate if the loss function did not show a significant decrease within 50 epochs. During the training phase, the data was fitted according to the above methods and parameters. Different learning rates and training / validation set ratios were used for training. Hyperparameters such as the learning rate, batch size, and number of training epochs were adjusted by examining the parameters. A small test set was used for prediction, and the parameters were gradually adjusted to obtain the desired training results.

[0048] Although the present invention has been described above with reference to embodiments, various modifications can be made and components can be replaced with equivalents without departing from the scope of the invention. In particular, as long as there is no structural conflict, the features in the disclosed embodiments can be combined with each other in any manner. The lack of an exhaustive description of these combinations in this specification is merely for the sake of brevity and resource conservation. Therefore, the present invention is not limited to the specific embodiments disclosed herein, but includes all technical solutions falling within the scope of the claims.

Claims

1. A power allocation method for the downlink of a deep learning-based cellular-free system, characterized in that, include: S1: First, consider the downlink of a large-scale multiple-input multiple-output (MLMI) network with L cells. Each cell includes a base station with M antennas and K users. A feedforward neural network with fully connected layers is used, consisting of a 2KL-dimensional input layer, N hidden layers, and a K+1-dimensional output layer, to fit the solution of all power allocation variables to the solution of a traditional algorithm. The relationship is established, and the training process is used to iterate the process step by step. S2: Construct a fully connected neural network structure model, transforming the problem between the power allocation variable set and the solution of the traditional algorithm into a problem between the weights W and the bias term b of the trained neural network. The aim is to optimize the loss function. minimize; S3: The dataset used by the neural network was generated from the DeepMIMO dataset. The optimal power allocation was obtained using traditional optimization methods. 80% of the samples were used for training, 10% for validation, and the remaining 10% formed the test set, independent of the training set. The specific operation method is as follows: a. First, read the data of the power allocation variable set and the corresponding solution of the traditional algorithm, and then perform dimensional transformation; b. Initialize the parameters required by the neural network; c. Define the network structure hierarchically, including the number of neurons and activation function parameters for each layer. At the model level, design a 6-layer fully connected network model. In this 6-layer fully connected network model, the ELU activation function is used between the first five layers, as shown in the following expression: , ; d. In the model merging stage, the Adam optimizer was used with an initial learning rate of 0.01 set in the function. The relative mean squared error was selected as the training metric, and an early stopping method was added with the patience parameter set to 50, meaning that training would be terminated if the loss function did not decrease significantly within 50 epochs. e. During the training phase, the data is fitted according to the above methods and parameters. Different learning rates and training set to validation set ratios are used for training. The parameters are checked to adjust hyperparameters such as learning rate, batch size, and number of training rounds. A test set with a small amount of data is used for prediction. The parameters are gradually adjusted to obtain training results that meet the requirements.

Images