CN113242066B - Multi-cell large-scale MIMO communication intelligent power distribution method - Google Patents
Multi-cell large-scale MIMO communication intelligent power distribution method Download PDFInfo
- Publication number
- CN113242066B CN113242066B CN202110505185.1A CN202110505185A CN113242066B CN 113242066 B CN113242066 B CN 113242066B CN 202110505185 A CN202110505185 A CN 202110505185A CN 113242066 B CN113242066 B CN 113242066B
- Authority
- CN
- China
- Prior art keywords
- user
- cell
- network
- power distribution
- users
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
Images
Classifications
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04B—TRANSMISSION
- H04B7/00—Radio transmission systems, i.e. using radiation field
- H04B7/02—Diversity systems; Multi-antenna system, i.e. transmission or reception using multiple antennas
- H04B7/04—Diversity systems; Multi-antenna system, i.e. transmission or reception using multiple antennas using two or more spaced independent antennas
- H04B7/0413—MIMO systems
- H04B7/0426—Power distribution
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/044—Recurrent networks, e.g. Hopfield networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/048—Activation functions
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04B—TRANSMISSION
- H04B7/00—Radio transmission systems, i.e. using radiation field
- H04B7/02—Diversity systems; Multi-antenna system, i.e. transmission or reception using multiple antennas
- H04B7/04—Diversity systems; Multi-antenna system, i.e. transmission or reception using multiple antennas using two or more spaced independent antennas
- H04B7/0413—MIMO systems
- H04B7/0456—Selection of precoding matrices or codebooks, e.g. using matrices antenna weighting
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04W—WIRELESS COMMUNICATION NETWORKS
- H04W72/00—Local resource management
- H04W72/04—Wireless resource allocation
- H04W72/044—Wireless resource allocation based on the type of the allocated resource
- H04W72/0473—Wireless resource allocation based on the type of the allocated resource the resource being transmission power
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04W—WIRELESS COMMUNICATION NETWORKS
- H04W72/00—Local resource management
- H04W72/50—Allocation or scheduling criteria for wireless resources
- H04W72/51—Allocation or scheduling criteria for wireless resources based on terminal or device properties
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04W—WIRELESS COMMUNICATION NETWORKS
- H04W72/00—Local resource management
- H04W72/50—Allocation or scheduling criteria for wireless resources
- H04W72/53—Allocation or scheduling criteria for wireless resources based on regulatory allocation policies
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04W—WIRELESS COMMUNICATION NETWORKS
- H04W72/00—Local resource management
- H04W72/50—Allocation or scheduling criteria for wireless resources
- H04W72/54—Allocation or scheduling criteria for wireless resources based on quality criteria
- H04W72/542—Allocation or scheduling criteria for wireless resources based on quality criteria using measured or perceived quality
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02D—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN INFORMATION AND COMMUNICATION TECHNOLOGIES [ICT], I.E. INFORMATION AND COMMUNICATION TECHNOLOGIES AIMING AT THE REDUCTION OF THEIR OWN ENERGY USE
- Y02D30/00—Reducing energy consumption in communication networks
- Y02D30/70—Reducing energy consumption in communication networks in wireless communication networks
Abstract
The invention discloses a multi-cell large-scale MIMO communication intelligent power distribution method, which comprises the steps of firstly generating a data set according to a traditional model-based optimal power distribution method, taking position information and channel state information of all users in a network as input characteristics, and taking the result of optimal power distribution as an output label, thereby generating a sample. The number of users contained in each sample of the data set is random. And further providing a seq2seq method based on a long-short term memory network, wherein a data set is used for training to obtain a mapping relation between the position information of the user and a power distribution strategy. The invention utilizes the communication network environment characteristics and the channel information to mine the relationship among the user characteristics, the power distribution and the system performance, designs the data-driven power distribution method, breaks through the limitation that the traditional method based on the neural network can only process the scene of fixed user number, and can flexibly adapt to the real-time change condition of the user number and the position only by training a single network.
Description
Technical Field
The invention relates to a multi-cell large-scale MIMO communication intelligent power distribution method, and belongs to the technical field of wireless communication.
Background
To meet the ever-increasing demands for communication equipment and data traffic, massive MIMO technology has been proposed and developed in the last few years. A base station in a large-scale MIMO network is provided with hundreds of antennas to obtain spatial multiplexing gain, and tens of users can be served on the same time-frequency resource, so that the frequency spectrum efficiency is effectively improved, and the base station becomes one of key technologies of 5G communication. The pilot pollution and the interference between users are usually the limitations of the MIMO system, and the interference is usually reduced and the spectrum efficiency is improved by reasonably allocating resources. And because a large-scale MIMO system is configured with a large-scale antenna array, a network can obtain larger spatial transmission gain, and a channel depends on large-scale fading, so that the problem of resource allocation is easier to solve.
Power allocation is a key technology for increasing the capacity of a communication system, and has been widely studied in academia. However, the traditional model-based method has the problem of high solving complexity. In view of the rapid development of data science and artificial intelligence, a communication system power control technology based on a neural network is developed. The method has the advantages that the full-connection neural network is used for fitting the traditional power distribution strategies such as the maximum minimum rate, the maximum rate, the rate and the like in the large-scale MIMO system by using the deep learning technology, a reasonable power distribution scheme can be provided under the condition that the position information of a user is known, and the performance of the method approaches to the traditional algorithm, and meanwhile, the complexity and the processing time of the optimization process are greatly reduced; there are also scholars who train a deep neural network to perform centralized power distribution in a de-cellular massive MIMO scenario based on dynamic cooperative clustering, or train a deep neural network through each access point, performing distributed power distribution using only locally available information as input. There have also been some studies on the decision to learn power allocation using the advantages of reinforcement learning. The methods fully utilize the capabilities of a neural network learning complex mode and approximation function mapping, and can greatly reduce the complexity and processing time of on-line execution while approximating the performance to the traditional algorithm. The limitation is that these methods are only for a fixed number of massive MIMO networks, i.e. both the number of cells (or access point number) and the number of users are fixed.
However, in practical communication scenarios, the number of users served at different times per cell is constantly changing, which is not in accordance with the assumptions made in previous studies. Although the deep neural network has flexibility and strong function fitting capability, it can only be applied to the problem that the input and target can be reasonably coded by vectors of fixed dimensions, which brings huge limitation to solving the wireless communication problem by using the neural network. At present, the relevant research is still insufficient, and a deep neural network based on residual error intensive connection of a convolutional neural network is designed by scholars for solving the problem and is used for processing the number of dynamically changed active users, so that the mapping of the large-scale fading coefficient and the power distribution strategy of the users is realized, the method limits the maximum number of users which can be served by each cell, and the large-scale fading coefficient of the number of inactive users is set to be 0. Thus, setting the number of the maximum service users in the fixed network to 0 inevitably causes a certain loss to the performance. From another point of view, the invention is inspired by successfully applying seq2seq neural network to process the problem of translation of variable-length text in the field of text processing, converts the problem of power distribution of the number of dynamic users into the problem of mapping from sequences of indefinite length to sequences, and learns the mapping relation by using the network of the seq2seq structure.
Disclosure of Invention
The technical problem is as follows: the invention aims to solve the defects that the existing power distribution method based on a model is high in solving complexity and only suitable for a fixed communication network scale, and provides a power distribution method based on a seq2seq neural network aiming at a scene of dynamic change of the number of multi-cell large-scale MIMO users. The method fully utilizes the relation between the user information and the power distribution mined by the neural network, can reduce the solving complexity while ensuring that the performance approaches the traditional method, and more importantly, can flexibly adapt to the scene of the change of the number and the position of the users in the communication network at any moment.
The technical scheme is as follows: in order to achieve the purpose, the invention adopts the technical scheme that: a multi-cell large-scale MIMO communication intelligent power distribution method. The technical scheme comprises the following steps:
the method comprises the following steps: and establishing a multi-cell large-scale MIMO system model with dynamically changed user number, and designing an objective function.
Step two: randomly setting the number and the position of users in each cell, solving an optimal power distribution strategy by using a max product SINR method, collecting user position information, channel state information and a corresponding optimal power distribution result, carrying out appropriate data preprocessing, taking the user position information and the channel state information in a network as input characteristics, taking the power distribution result as an output label, and collecting for multiple times to form a data set.
Step three: and constructing a seq2seq network model for processing variable-length sequences of input and output.
Step four: and (4) training the data set generated in the step two by using the network built in the step three.
Step five: and inputting a sequence generated by the user position and the channel state information in the current network into the trained seq2seq network model to obtain an optimal power distribution sequence in the current scene, and performing inverse normalization processing to obtain an optimal power distribution result.
In the first step, a multi-cell large-scale MIMO system model with dynamically changed user number is established. Suppose a massive MIMO communication network has L cells, each of which has a base station equipped with M antennas. Assuming that the number of users in each cell changes dynamically, the number of users served by the ith cell is k i The number of users served by each cell in the network can be expressed as K ═ K i I 1, a, L, with a i ={1,2,...,k i Represents the user number set in the ith cell, and the size of the set is k i Determining that N is k 1 +k 2 +...+k L Is the total number of all users in the current system.
Channel usage between user n and base station j in cell lTo indicate that the base station of cell l transmits downlink signalsWhereinIs a signal sent to user n in cell l, using a precoding vectorTo control the spatial directionality of the transmission, the precoding vector satisfies | | w l,n || 2 =1,ρ l,n Indicating the transmission power allocated by the base station to user n in cell i. Further, the downlink spectrum efficiency of the user n in the cell l of the massive MIMO system is:
wherein
Representing the ratio, σ, of data for downlink transmission in each coherent block 2 Representing the noise power, the superscript dl for SINR represents the downlink, E { · } represents the expectation, | · | represents the absolute value operation,denotes w l,n The conjugate transpose of (c).
For convenience of representation, the above spectral efficiency is re-expressed as:
wherein
a l,n And b i,k,l,n Mean channel gain and mean interference gain are indicated separately.
The objective function considered in step one is:
And step two, generating a data set. In each data acquisition process, the number k of users in each cell is randomly set in a network area l And location x of each user l,n Then calculating large scale fading coefficientAnd channel correlation matrixObtaining estimated channel vector by using channel estimation method based on minimum mean square error algorithmComputing a precoding vector w l,n . Averaging the estimated channels by Monte Carlo to obtain { a l,n And { b }and i,k,l,n }. Power allocation strategy for maximizing objective function by traditional geometric programming solution
Definition S ═ (S) 1 ,s 2 ,...,s N ) Is an ordered sequence of numbers in which s i The vector (G, D, a, I) is a one-dimensional vector for representing information about the user I in the system, and G ═ G 1 ,g 2 ,...,g L ) Denotes the channel gain between the user and L base stations, D ═ D 1 ,d 2 ) Representing usersPosition information, where a denotes a signal amount transmitted by a user, and I denotes an amount of (I) transmitted by a user 1 ,i 2 ,...,i N-1 ) The amount of interference to that user for other users in the system. The information of the N users together form an ordered sequence S. Taking S as an input feature, ρ * As an output tag. One sample is formed for one data acquisition, resulting in a data set with a large number of samples. To better fit the network, a logarithmic transformation and normalization process is used on the samples.
The size of I is fixed to the maximum number of users which can be served by the network, and the I of each sample is supplemented to the fixed length by adopting a processing mode of filling 0 at the tail end.
And in the third step, a network model is built, and the built seq2seq model is divided into two parts, namely an encoder part and a decoder part.
In order to avoid the limitation that the traditional Recurrent Neural Network (RNN) can have long-term dependence, the encoder part uses a long-term memory network (LSTM) to fully mine the correlation of users at different positions in the communication network. The input dimension of the encoder is set to the dimension of the information of the single user described in step two, i.e. s i Of (c) is measured. The encoder reads the sequences in sequence in the order of the sequence S, thereby converting the sequences of indefinite length into intermediate vectors of fixed dimensions. The processing of the encoder can be expressed as:
wherein x is t Is an input variable at time t, W is a weight matrix of the neural network unit, b represents bias, σ (-) and g (-) represent sigmoid function and tanh function, respectively, f t 、i t 、C t 、And o t All are intermediate variables in the process, h t For the output variable of LSTM, i.e. the hidden state obtained after the input information passes through the encoder of the seq2seq network, assuming that the input of the sequence is completed at time t equal to N, then h N All input information is processed and converted into a context vector with fixed length.
The decoder part is another LSTM network and needs to ensure that its hidden layer dimension is the same as the dimension of the context vector. And then connecting a full connection layer, converting the output of each step into a value, and performing inverse normalization processing on the value to obtain the power value which is calculated by the seq2seq model and is distributed to the current user.
Before training the established neural network, firstly, initializing parameters of the model, and setting the batch size n batch Training iteration number n epoch . Each time n is input batch The characteristic part S of each sample is processed by an encoder to obtain an intermediate variable h N Inputting the input data into a decoder end, setting a guiding strength factor, selecting the output of the previous step as the current input variable with a certain probability in each step, or else, using the corresponding label value as the input variable, and then obtaining the output vector of the modelCalculating a loss functionAnd updating model parameters by adopting an adaptive moment estimation optimization method. And freezing the network parameters after the training is finished.
The algorithm application process in the fifth step is specifically as follows: and preprocessing the user position and the channel state information in the current network to generate a sequence with a fixed format, inputting the sequence into a trained neural network model to obtain an optimal power distribution sequence in the current scene, and performing inverse normalization processing on the optimal power distribution sequence to obtain an optimal power distribution result.
Has the beneficial effects that: compared with the prior art, the technical scheme of the invention has the following beneficial technical effects:
the invention provides a multi-cell large-scale MIMO communication intelligent power distribution method, which considers the situation that the number and the position of users change in real time in an actual scene. The long-time and short-time memory network is utilized to mine the correlation of users at different positions, and the mapping relation between the user position information and the channel state information and the power distribution is fitted, so that the performance can approach the traditional algorithm, and the method can adapt to the scene that the number and the position of the users are constantly changed in real time.
Drawings
Fig. 1 is a flowchart of an intelligent power allocation method for multi-cell large-scale MIMO communication according to the present invention.
Fig. 2 is a structural diagram of a seq2seq neural network constructed in the embodiment of the present invention.
Fig. 3 and 4 are graphs comparing the performance of an intelligent power allocation method for multi-cell massive MIMO communication according to an embodiment of the present invention with that of a conventional method.
Detailed Description
The present invention will now be described in further detail with reference to the drawings and detailed description of exemplary embodiments of the invention, it being understood that the embodiments described are merely illustrative of some, and not restrictive, of the invention. All other embodiments, which can be obtained by a person skilled in the art without making any creative effort based on the embodiments in the present invention, belong to the protection scope of the present invention.
The following is an implementation manner of the intelligent power allocation method suitable for multi-cell large-scale MIMO communication of the present invention:
the method comprises the following steps: and establishing a multi-cell large-scale MIMO system model with dynamically changed user number, and designing an objective function.
Step two: the number and the position of users in each cell are randomly set, an optimal power distribution strategy is solved by using a traditional optimization method, user position information, channel state information and a corresponding optimal power distribution result are collected, appropriate data preprocessing is carried out, the user position information and the channel state information in a network are used as input characteristics, the power distribution result is used as an output label, and a data set is formed by multiple times of collection.
Step three: and constructing a seq2seq network model for processing variable-length sequences of input and output.
Step four: and (4) training the data set generated in the step two by using the network built in the step three.
Step five: and inputting a sequence generated by the user position and the channel state information in the current network into the trained seq2seq network model to obtain an optimal power distribution sequence in the current scene, and performing inverse normalization processing to obtain an optimal power distribution result.
In the first step, a multi-cell large-scale MIMO system model with dynamically changed user number is established. The embodiment of the invention assumes that a large-scale MIMO communication network is provided with 4 cells, the coverage area of each cell is 0.5km multiplied by 0.5km, and the center of each cell is provided with a base station which is provided with 100 antennas. Assuming that the number of users in each cell is not fixed, the number of users served by the ith cell is k i The number of users served in each cell in the network is represented as K ═ K i 1.. times.l., using a, i ═ 1 i ={1,2,...,k i Denotes the set of user numbers in the ith cell, the size of the set is denoted by k i Determining that N is k 1 +k 2 +...+k L Is the total number of all users in the current system. In the embodiment of the invention, the user set K is set to be 0,5]Random permutation of integers of intervals. Each user is located more than 35m from the base station. Assuming a communication bandwidth of 10MHz, the total receiver noise power σ 2 Set to-94 dBm. The pilot multiplexing factor is 1. The uplink transmit power ρ of each user equipment is 20 dBm. The number of monte carlo simulations was set to 100.
Channel usage between user n and base station j in cell lIt is shown that in the present embodiment, it is assumed that it obeys the rayleigh fading distribution:
wherein the content of the first and second substances,is a space correlation matrix obtained by the base station side, and the normalized trace of the matrixThe large-scale fading coefficient simulating geometric path loss and shadow fading represents the average channel gain between one antenna of the base station j and the user n of the cell l, and is modeled as:
where γ -148dB represents the median of the channel gain at a reference distance of 1km between the user and the base station, α -3.76 is the path loss coefficient,is the distance between the user and the base station on a two-dimensional plane.
The channel estimation adopts a standard minimum mean square error channel estimation method, and the estimation of the channel between the base station j and the user n in the cell l can be:
wherein, the first and the second end of the pipe are connected with each other,is a noise, and is a sound wave,set P n ={l′∈{1,...,L}:n≤k l′ Denotes the index set of all cells containing the user number n, the error of channel estimationIndependently of
Let the base station of cell l transmit downlink signalsWhereinIs a signal sent to user n in cell l, using a precoding vectorTo control the spatial directionality of the transmission, the precoding vector satisfies | | w l,n || 2 =1,ρ l,n Indicating the transmission power. The following precoding scheme is used in the embodiments of the present invention:
wherein the content of the first and second substances,
then, the achievable downlink spectrum efficiency of the user n in the cell l of the massive MIMO system is:
wherein
The spectral efficiency is re-expressed as:
wherein
a l,n And b i,k,l,n Respectively, mean channel gain and mean interference gain.
The objective function of the first step is designed as follows:
and step two, generating a data set. When generating a sample, the number k of users in each cell needs to be randomly set in the network area l And location of individual usersThen calculating large scale fading coefficientAnd channel correlation matrixObtaining an estimated channel vector using a minimum mean square error based channel estimation methodAccording to the aboveCalculating a precoding vector w according to the precoding scheme l,n . Averaging the estimated channels by Monte Carlo to obtain { a l,n And { b }and i,k,l,n }. Power allocation strategy for maximizing objective function by traditional geometric programming solutionThis method of solving the optimal solution of power allocation by geometric programming can be called max product SINR method.
Definition S ═ (S) 1 ,s 2 ,...,s N ) Is an ordered sequence of numbers in which s i The vector (G, D, a, I) is a one-dimensional vector for representing information about the user I in the system, and G ═ G 1 ,g 2 ,g 3 ,g 4 ) Denotes the channel gain between the user and L base stations, D ═ D 1 ,d 2 ) Indicating the location information of the user, a ═ a indicating the amount of signal transmitted by the user, and I ═ I indicating the amount of signal transmitted by the user 1 ,i 2 ,...,i N-1 ) The amount of interference to that user for other users in the system. The information of the N users together form an ordered sequence S. It should be noted that the size of each sample is different since the total number of users at a time varies. But each s in the input features due to network limitations i Must be fixed, s i The size of G, D, A is fixed, only the size of I varies with N, so the size of I is fixed as the maximum number of users that can be served by the network, in this embodiment, since the user set K is set to [0,5 ]]The integers of the interval are randomly arranged, so that the maximum number of users in the network is 14, and the I of each sample is supplemented to 14 values by adopting a processing mode of filling 0 at the tail. Then s i Is fixed at 20.
Taking S as an input feature, ρ * As an output tag. A random scatter of the user at a time generates one sample, ultimately generating a data set with a large number of samples. To better fit the network, a logarithmic conversion and normalization process is used on the samples. 200000 samples were generated.
And (3) building a network model in the third step, wherein the built seq2seq model is divided into two parts, namely an encoder part and a decoder part.
In order to avoid the limitation of long-term dependence of the traditional Recurrent Neural Network (RNN), the encoder part uses a long-term memory network (LSTM) network to fully mine the relevance of users at different positions in the communication network. The input dimension of the encoder is set to the dimension of the information of the single user described in step two, i.e. s i Dimension of (c), 20. The number of neurons in each hidden layer is 128. The encoder reads the sequence sequentially in the order of the sequence S, thereby converting the indefinite length sequence into a 128-dimensional hidden state. h is N Is an intermediate variable with fixed length which is converted by processing all input information.
The decoder part is another LSTM network and needs to guarantee that its hidden layer dimension and the dimension of the intermediate variables are the same. And then connecting a full connection layer, converting the output of each step into a value, and performing inverse normalization processing on the value to obtain the power value which is calculated by the seq2seq model and is distributed to the current user.
The model training process in the step four specifically comprises the following steps: firstly, initializing the parameters of the model, and setting the batch size n batch 8, training iteration number n epoch 40. Each time n is input batch The characteristic S of each sample is processed by an encoder to obtain an intermediate variable h N Inputting the input data into a decoder end, setting a guiding strength factor, selecting the output of the previous step as the current input variable with a certain probability in each step, or else, using the corresponding label value as the input variable, and then obtaining the output vector of the modelCalculating a loss functionAnd updating model parameters by adopting an adaptive moment estimation optimization method. And freezing the network parameters after the training is finished.
The algorithm application process in the step five specifically comprises the following steps: and preprocessing the user position and the channel state information in the current network to generate a sequence with a fixed format, inputting the sequence into a trained neural network model to obtain an optimal power distribution sequence in the current scene, and performing inverse normalization processing on the optimal power distribution sequence to obtain an optimal power distribution result.
The invention evaluates the multi-cell large-scale MIMO communication intelligent power distribution method provided by the invention through simulation experiments. After a seq2seq network model is trained and frozen, power allocation is performed by respectively using an average allocation method, a max product SINR allocation method and the intelligent allocation method based on the seq2seq network, and fig. 3 is a cumulative probability distribution (CDF) curve of user average Spectral Efficiency (SE) obtained by using the three methods. It can be found that the average allocation method obviously has the worst performance, and the trained seq2seq power allocation model is very close to the max product SINR power allocation method.
The samples in the test set contain a variety of user numbers. Fig. 4 compares CDF curves of the system and the rate obtained by performing power allocation respectively by the method of the present patent and the max product SINR method when there are 10, 11, 12, 13, and 14 users in the system respectively. The power distribution model based on the seq2seq network can be well adapted to the condition of the change of the number of users, the calculation complexity is greatly reduced, and meanwhile, the system and the speed performance can be ensured to approach the traditional method under the scene of various numbers of users.
Those skilled in the art can adaptively change the modules in the embodiment and set them in an optimization method or apparatus different from the embodiment. Specifically, a plurality of modules in the embodiment may be combined into one module, or one module may be divided into a plurality of sub-modules, which is applied to a method or an apparatus of the same technical idea as the embodiment.
Claims (6)
1. A multi-cell large-scale MIMO communication intelligent power distribution method is characterized by comprising the following steps:
the method comprises the following steps: establishing a multi-cell large-scale MIMO system model with dynamically changed user number, and designing a target function;
step two: randomly setting the number and the position of users in each cell, solving an optimal power distribution strategy by using a max product SINR method, collecting user position information, channel state information and corresponding optimal power distribution results, carrying out data preprocessing, taking the user position information and the channel state information in a network as input characteristics, taking the power distribution results as output labels, and collecting for multiple times to form a data set;
step three: building a seq2seq network model for processing variable-length sequences of input and output, wherein the seq2seq represents a sequence-to-sequence;
step four: training the data set generated in the step two by using the seq2seq network model built in the step three;
step five: inputting a sequence generated by a user position and channel state information in a current network into a trained seq2seq network model to obtain an optimal power distribution sequence in a current scene, and performing inverse normalization processing to obtain an optimal power distribution result;
the concrete process of constructing the seq2seq network model in the third step is as follows:
the constructed seq2seq model is divided into two parts: an encoder section and a decoder section;
the encoder part uses a long-term memory network to mine the correlation of users at different positions in the communication network; the encoder sequentially reads the sequences according to the sequence order, thereby converting the sequences with indefinite length into intermediate vectors with fixed dimensionality; the processing of the encoder is represented as:
wherein x t Is an input variable at time t, W is a weight matrix of the neural network, b represents bias, σ (-) and g (-) represent sigmoid function and tanh function, respectively, f t 、i t 、C t 、And o t All are intermediate variables in the process, h t Assuming that the input of the sequence is completed at time t N, h is the hidden state obtained by the LSTM output variable at time t, i.e. the input information after passing through the encoder of the seq2seq network N All input information is processed and converted into a context vector with fixed length; subscript f corresponds to forget gate, W f Weight matrix representing forgetting gate, b f Representing a forgetting gate bias parameter, i representing an input gate, c representing a state unit, and o representing an output gate; LSTM represents a long-and-short term memory network;
the decoder part is another LSTM network whose hidden layer dimension is the same as that of the context vector; the LSTM network is connected with a full connection layer, the output of each step is converted into a value, and the value is subjected to reverse normalization processing to obtain a power value which is calculated by a seq2seq model and is allocated to a current user;
the seq2seq network model training process in the fourth step specifically comprises the following steps:
firstly, initializing parameters of a seq2seq network model, and setting a batch size n batch And training iteration number n epoch (ii) a Each time n is input batch The characteristic S of each sample is processed by an encoder to obtain an intermediate variable h N Inputting the input data into a decoder end, setting a guiding strength factor, selecting the output of the previous step as the current input variable with a certain probability in each step, or else, using the corresponding label value as the input variable, and then obtaining the output vector of the modelCalculating a loss function ρ * Updating model parameters by adopting an adaptive moment estimation optimization method for a power distribution strategy for maximizing the objective function; and freezing the network parameters after the training is finished.
2. The intelligent power allocation method for multi-cell massive MIMO communication according to claim 1, wherein the specific step of establishing the multi-cell massive MIMO system model with dynamically changing user number in step one comprises:
the large-scale MIMO communication network is assumed to have L cells, each cell is provided with a base station and M antennas; assuming that the number of users in each cell changes dynamically, the number of users served by the ith cell is k i The number of users served in each cell in the network is represented as K ═ K i I 1, a, L, with a i ={1,2,...,k i Denotes the set of user numbers in the ith cell, the size of the set is denoted by k i Determining N ═ k 1 +k 2 +...+k L The total number of all users in the current system;
channel usage between user n and base station j in cell lTo indicate that the base station of cell l transmits downlink signalsWhereinIs a signal sent to user n in cell l, using a precoding vectorTo control the spatial directivity of the transmission, the precoding vector is satisfiedρ l,n Represents the transmission power allocated by the base station to user n in cell l; the downlink spectrum efficiency of the user n in the cell l is as follows:
wherein
Representing the ratio, σ, of data used for downlink transmission in each coherent block 2 Representing the noise power, the superscript dl of the SINR represents the downlink, E {. cndot. } represents the expectation, l | - | represents the absolute value operation,denotes w l,n The ranges of L and j are [1, L ]],Is a complex set of dimensions M, p i,k Represents the transmission power allocated by the base station to user k in cell i,representing the channel between user n and base station i in cell i.
5. The intelligent power allocation method for multi-cell massive MIMO communication according to claim 1, wherein the specific steps of the second step comprise:
in each data acquisition process, the number k of users in each cell is randomly set in a network area l And location x of each user l,n Then calculating large scale fading coefficientAnd channel correlation matrixObtaining estimated channel vector by using channel estimation method based on minimum mean square error algorithmComputing a precoding vector w l,n (ii) a Averaging the estimated channels by Monte Carlo to obtain { a } l,n And { b }and i,k,l,n }; power allocation strategy for maximizing objective function by traditional geometric programming solution
Definition S ═ (S) 1 ,s 2 ,...,s N ) Is an ordered sequence of numbers in which s i The vector (G, D, a, I) is a one-dimensional vector for representing information about the user I in the system, and G ═ G 1 ,g 2 ,...,g L ) Denotes the channel gain between the user and L base stations, D ═ D 1 ,d 2 ) Indicates the location information of the user, a ═ a) indicates the amount of signals transmitted by the user, and I ═ I indicates the amount of signals transmitted by the user 1 ,i 2 ,...,i N-1 ) The interference amount of other users in the system to the user; the information of N users jointly form an ordered sequence S; taking S as an input feature, ρ * As an output tag; forming a sample by primary data acquisition, and carrying out logarithmic transformation and standardization treatment on the sample to finally generate a data set;
subscript l denotes the l cell, subscript n denotes the n user, superscript j denotes the j base station, a l,n And b i,k,l,n Respectively representing the average channel gain and the average interference gain, s i Wherein i has a range of [1, N ]]。
6. The intelligent power allocation method for multi-cell massive MIMO communication as claimed in claim 5, wherein the size of I is fixed to the maximum number of users that the network can serve, and the I of each sample is supplemented to a fixed length by a processing method of filling 0 at the end.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202110505185.1A CN113242066B (en) | 2021-05-10 | 2021-05-10 | Multi-cell large-scale MIMO communication intelligent power distribution method |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202110505185.1A CN113242066B (en) | 2021-05-10 | 2021-05-10 | Multi-cell large-scale MIMO communication intelligent power distribution method |
Publications (2)
Publication Number | Publication Date |
---|---|
CN113242066A CN113242066A (en) | 2021-08-10 |
CN113242066B true CN113242066B (en) | 2022-07-26 |
Family
ID=77133174
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202110505185.1A Active CN113242066B (en) | 2021-05-10 | 2021-05-10 | Multi-cell large-scale MIMO communication intelligent power distribution method |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN113242066B (en) |
Families Citing this family (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN114826833B (en) * | 2022-03-11 | 2024-01-26 | 国网福建省电力有限公司经济技术研究院 | Communication optimization method and terminal for CF-mMIMO in IRS auxiliary MEC |
CN114915322B (en) * | 2022-05-06 | 2024-03-08 | 东南大学 | Multi-cell large-scale MIMO communication power distribution method based on semi-supervised learning and feature augmentation |
Citations (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN112153615A (en) * | 2020-09-15 | 2020-12-29 | 南京信息工程大学滨江学院 | Deep learning-based user association method in multi-cell cellular D2D equipment |
Family Cites Families (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN106899993B (en) * | 2017-04-02 | 2020-10-09 | 上海无线通信研究中心 | Network optimization method facing large-scale MIMO network and base station thereof |
CN111901862B (en) * | 2020-07-07 | 2021-08-13 | 西安交通大学 | User clustering and power distribution method, device and medium based on deep Q network |
CN112153617B (en) * | 2020-09-15 | 2022-07-12 | 南京信息工程大学滨江学院 | Terminal equipment transmission power control method based on integrated neural network |
CN112492691B (en) * | 2020-11-26 | 2024-03-26 | 辽宁工程技术大学 | Downlink NOMA power distribution method of depth deterministic strategy gradient |
-
2021
- 2021-05-10 CN CN202110505185.1A patent/CN113242066B/en active Active
Patent Citations (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN112153615A (en) * | 2020-09-15 | 2020-12-29 | 南京信息工程大学滨江学院 | Deep learning-based user association method in multi-cell cellular D2D equipment |
Also Published As
Publication number | Publication date |
---|---|
CN113242066A (en) | 2021-08-10 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
Lee et al. | Deep power control: Transmit power control scheme based on convolutional neural network | |
CN110401964B (en) | Power control method based on deep learning for user-oriented center network | |
CN113242066B (en) | Multi-cell large-scale MIMO communication intelligent power distribution method | |
CN112153616B (en) | Power control method in millimeter wave communication system based on deep learning | |
CN113438002B (en) | LSTM-based analog beam switching method, device, equipment and medium | |
CN113596785A (en) | D2D-NOMA communication system resource allocation method based on deep Q network | |
CN113411817B (en) | Wireless system interference neural network prediction method based on wireless interference model | |
CN115102592B (en) | Multi-user MIMO beam forming method based on federal learning | |
CN114885340B (en) | Ultra-dense wireless network power distribution method based on deep migration learning | |
CN113239632A (en) | Wireless performance prediction method and device, electronic equipment and storage medium | |
CN113381790A (en) | AI-based environment knowledge assisted wireless channel feedback method | |
CN113038612B (en) | Cognitive radio power control method based on deep learning | |
CN117240331A (en) | No-cellular network downlink precoding design method based on graph neural network | |
Liu et al. | Power allocation in ultra-dense networks through deep deterministic policy gradient | |
CN113595609B (en) | Collaborative signal transmission method of cellular mobile communication system based on reinforcement learning | |
CN112243283B (en) | Cell-Free Massive MIMO network clustering calculation method based on successful transmission probability | |
CN114710187A (en) | Power distribution method for multi-cell large-scale MIMO intelligent communication under dynamic user number change scene | |
CN114268348A (en) | Honeycomb-free large-scale MIMO power distribution method based on deep reinforcement learning | |
CN107835043A (en) | A kind of method of information transfer outage probability in rapid evaluation wireless power communication | |
CN108173580B (en) | Quantum social emotion optimization-based Massive MIMO uplink system power distribution method | |
CN113473580B (en) | User association joint power distribution method based on deep learning in heterogeneous network | |
CN109194371B (en) | Low-complexity power distribution method of distributed MISO system based on beam forming | |
CN113114314B (en) | Large-scale MIMO performance improving method integrating hybrid enhancement and self-encoder | |
CN111914923B (en) | Target distributed identification method based on clustering feature extraction | |
CN114650086B (en) | Cross-frequency-band communication beam prediction method assisted by deep learning |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
GR01 | Patent grant | ||
GR01 | Patent grant |