High dynamic traffic and edge user number prediction method based on multivariate data fusion

Abstract

The invention discloses a high-dynamic traffic and edge user number prediction method based on multivariate data fusion. The method comprises the following steps: acquiring time sequence data of historical RRH traffic and edge user number; converting the sequence data of the historical RRH flow and the edge user number into embedded matrix features; and predicting future RRH flow and the number of edge users based on the embedded matrix characteristics. According to the invention, through fusion of flow and people number historical sequence data, related features are learned, and time and space features in the sequence data are extracted and added into a network structure for training. Multiple groups of base stations can be processed at the same time, a model is helped to mine hidden associations from different dimensions, and more accurate prediction is achieved; the high-dynamic traffic and people number prediction framework can be flexibly expanded to cope with different network scenes; through RRH flow and edge user number prediction, the network can realize more intelligent resource allocation, and the overall efficiency and user experience of the network are improved.

Description

Technical Field

[0001] The present invention relates to the field of edge mobility, and in particular to a method for predicting high-dynamic traffic and the number of edge users based on multivariate data fusion. Background Art

[0002] In recent years, to address the rapidly growing number of edge mobile users and the surging data traffic demand, China Mobile Research Institute has proposed the Cloud Radio Access Network (C-RAN) concept. This concept includes distributed Remote Radio Heads (RRHs) and a centralized Baseband Unit Pool (BBU). To fully leverage the power of the C-RAN architecture, a key challenge is designing an appropriate matching solution between the RRHs and BBUs to minimize overall network costs while ensuring user quality of service.

[0003] However, designing a dynamic, real-time RRH-BBU matching solution is challenging because future RRH connection status is difficult to obtain. Given the significant and highly dynamic nature of RRH traffic and user usage, the RRH-BBU matching solution at each moment must be determined based on the current RRH connection status. The current matching solution cannot be directly applied to future moments. Therefore, to implement a dynamic, real-time RRH-BBU matching solution, it is necessary to obtain the RRH connection status at future moments—in other words, to predict the number of connected RRH users and traffic at the next point in time. However, achieving this prediction for multiple RRHs is not easy. First, due to the high mobility and dynamic nature of users, traffic and user usage between RRHs vary significantly across time and space. For example, RRHs located in commercial areas generate more traffic during the day. Second, due to the lack of RRH location information in the data, direct spatial feature extraction is impossible. Finally, RRH traffic and the number of edge users exhibit abrupt changes, and the amount of data available at a single point in time is limited. Therefore, it is necessary to extract useful information from a wider timeframe. Therefore, an effective method is needed to accurately predict highly dynamic traffic and edge user numbers. Summary of the Invention

[0004] The purpose of the present invention is to provide a highly dynamic traffic and edge user number prediction method based on multivariate data fusion, which is used to solve the problems of improper resource allocation, poor network performance and user experience caused by the lack of real-time information in the RRH-BBU matching solution.

[0005] A highly dynamic traffic and edge user number prediction method based on multivariate data fusion includes the following steps:

[0006] Step 1: Obtain time series data of historical RRH traffic and edge user numbers;

[0007] Step 2: Convert the historical RRH traffic and edge user number sequence data into embedding matrix features;

[0008] Step 3: Predict future RRH traffic and number of edge users based on the embedded matrix features.

[0009] Step 1 specifically includes:

[0010] Acquires multiple heterogeneous data including RRH connection data, user attribute data, and traffic usage data, and stores them in traffic tables and user attribute tables;

[0011] Extract all RRH traffic sizes and the number of connected users in one hour from the traffic table and user attribute table in hourly fine-grained manner to obtain the time series tensor X t :

[0012]

[0013] Where t represents the tth hour, n represents the total number of n RRHs, traffic represents the traffic volume of an RRH during this time period, and user represents the number of edge users of an RRH.

[0014] For the time series tensor X of RRH traffic and edge user number t , extracting the time order sequence tensor L that expresses the time order characteristics t ;

[0015] For the time series tensor X of RRH traffic and edge user number t , on the premise of hourly fine-grained extraction, considering the mean, variance and median of all RRH traffic and edge user numbers within an hour, the change trend sequence tensor N of the overall change trend of RRH traffic and edge user numbers between each time period is obtained t :

[0016]

[0017] in, Represent the mean, variance and median of RRH traffic and edge user number respectively;

[0018] Multiply each tensor by the number of time periods to form the sequence data of RRH traffic and edge user number:

[0019] Given a time series tensor X containing P time periods t , obtain spatiotemporal series data X t-P:t =P×X t ;

[0020] Given a time sequence tensor L containing P time periodst ,, obtain time sequence data L t-P:t =P×L t ;

[0021] Given a trend sequence tensor N containing P time periods t , obtain the change trend sequence data N t-P:t =P×N t .

[0022] Step 2 specifically includes:

[0023] The embedding layer is constructed by the formula:

[0024] E t =CNN embedding (X t-P:t ,L t-P:t ,N t-P:t )

[0025] The sequence data X t-P:t , L t-P:t and N t-P:t Convert to embedding matrix E t ∈R P×N ,CNN(·) represents the convolutional layer, where P represents P time periods starting from time period t, and N represents N RRHs;

[0026] Two trainable randomly initialized matrices S and T are introduced to represent the time series data X t-P:t The spatial and temporal features of t ∈R P×N And the spatial feature matrix S and the temporal feature matrix T are expressed by the formula

[0027] M t =Concat(E t ,S t ,T t )

[0028] Combine and generate mixed features M t The spatial feature matrix S contains the latitude and longitude information of the RRHs in the time series of different RRH traffic and edge user numbers. In the temporal feature matrix T, the same time points between the time series data of different RRH traffic and edge user numbers share the same eigenvalue.

[0029] We will get the feature M t , convolution operation is performed through K convolutional layers with residual connections;

[0030] According to the formula:

[0031]

[0032] The output layer generates the corresponding embedding matrix feature Z of RRH traffic and edge user number t ∈R P×N ;

[0033] Wherein, P represents P time periods starting from time period t, and N represents N RRHs.

[0034] We will get the feature M t , the convolution operation is performed through K convolutional layers with residual connections, specifically:

[0035] Each time a convolutional network is passed, the residual connection will superimpose the output of the previous convolutional network to the input of the next convolutional network; the calculation method of the k-th convolutional layer is:

[0036]

[0037] Step 3 specifically includes:

[0038] Embed the RRH traffic and edge user number into matrix features Z t Divide into different small time segments:

[0039] By dividing each variable into time segments with a step size of s, the input embedding matrix feature Z t ∈R P×N Convert to Z t ∈R C×L×N In the form of, where L represents the length of each time segment, C is the number of time segments divided in a variable, and the calculation method is

[0040] The divided Z t ∈R C×L×N Rearrange to Z t ∈R N×C×L ;

[0041] The rearranged three-dimensional feature tensor Z t ∈R N×C×L Get the feature tensor P t ∈R N×C×D , where D represents the target dimension of the feature tensor;

[0042] The feature tensor P is extracted by stacking multiple single-layer hybrid MLP layers. t ∈R N×C×D characteristics and predict future RRH traffic and edge user numbers.

[0043] The rearranged three-dimensional feature tensor Z t ∈R N×C×l The characteristic tensor P t ∈RN×C×D , specifically:

[0044] The rearranged three-dimensional feature tensor Z t ∈R N×C×L , through a fully connected layer, the high-dimensional feature tensor of each sample is matched to a low-dimensional feature tensor to obtain the feature tensor P t ∈R N×C×D .

[0045] The feature tensor P is extracted by stacking multiple single-layer hybrid MLP layers. t ∈R N×C×D The characteristics of the RRH are used to predict the future RRH traffic and number of edge users, specifically:

[0046] A single-layer hybrid MLP architecture is used to extract relevant features across time segments and hidden features within a time segment respectively;

[0047] The extracted correlation features across time segments and hidden features within time segments are respectively stacked into a multi-layer hybrid MLP using the residual connection step;

[0048] Through multiple layers of mixed MLP layers, the output feature tensor P is obtained t ∈R N×C×D ;

[0049] Through an MLP network, the output feature tensor P t ∈R N×C×D Perform restoration to predict future RRH traffic and the number of edge users.

[0050] A single-layer hybrid MLP architecture is used to extract relevant features across time segments, specifically:

[0051] For the feature tensor P t Perform normalization processing;

[0052] The normalized feature tensor is used as the input of the self-attention mechanism. The attention weight matrix between different time segments is obtained through the self-attention mechanism to measure the feature importance between time segments. The weight matrix is ​​multiplied by the normalized feature tensor and activated by the Softmax function. The activated result is passed through the Linear layer to obtain the output of the self-attention mechanism network.

[0053] Set the weight matrix size of the shared MLP network to W M ∈R C×C , the shared MLP network shares parameters between different variables and different features within the time slice, W M Learn the correlation between different time segments through linear transformation;

[0054] At the same time, the normalized feature tensor P t ∈R N×C×D Transformed into feature tensor P by matrix transposition t ∈R N×D×C , as the input of the MLP network, input feature P t is converted to R N×D×C , where C is the number of time segments, and the weight matrix acts on the last dimension of the time segment dimension, matching the input dimension;

[0055] The output of the shared MLP network is processed by the gated attention mechanism unit, and a weight matrix W is learned through the gated attention unit. G ∈R N×C×D To amplify important features and suppress irrelevant features, the weight matrix is ​​calculated as follows:

[0056] W G =softmax(FC(P t ))

[0057] Where FC(·) represents the linear layer;

[0058] The output of the gated attention unit is obtained by adding the attention weights W G and the hidden tensor P output from the previous layer of the network t Dot product gives:

[0059]

[0060] Convert the output of the gated attention mechanism unit to the feature tensor P t ∈R N×D×C Transform to P by matrix transposition t ∈R N×C×D ;

[0061] The result of the conversion P t ∈R N×C×D Add to the output of the self-attention network and normalize;

[0062] The normalized result is residually connected to the input of the self-attention mechanism. Each time a layer of the network is passed, the residual connection will superimpose the output of the previous layer of the network on the input of the next layer of the network, and finally output the relevant features across time segments.

[0063] A single-layer hybrid MLP architecture is used to extract the hidden feature correlation within the time segment, specifically:

[0064] For the correlation features P across time segments t ∈R N×C×D Perform normalization processing;

[0065] The normalized results are passed through a shared MLP unit, which learns the correlation between different embedded features within a time slice and shares parameters across different variables and time slices. The weight matrix size of the network is B×B.

[0066] The output of the shared MLP network is processed by the gated attention mechanism unit and then residually connected. Each time it passes through a layer of the network, the residual connection will superimpose the output of the previous layer of the network on the input of the next layer of the network; finally, the hidden features within the time segment are output.

[0067] Through an MLP network, the output feature tensor P t ∈R N×C×D Perform restoration and predict future RRH traffic and edge user numbers, specifically:

[0068] The output feature tensor P t ∈R N×C×D Converted into a two-dimensional form P by unfolding operation t ∈R N×(DC) ;

[0069] After expansion, P t ∈R N×(DC) Through a linear layer processing, the matching function from the current feature space to the prediction space is learned. The output dimension of the linear layer is set to N×F to directly correspond to the dimension of the target prediction value;

[0070] The linear layer outputs an intermediate feature representation P t ∈R N×F ;

[0071] The output of the linear layer P t ∈R N×F Transpose to obtain the final RRH traffic and edge user number prediction result Y∈R F ×N .

[0072] This paper proposes a method for traffic and headcount analysis based on multivariate data fusion. By fusing historical sequence data of traffic and headcount, the present invention learns relevant features, extracts temporal and spatial features from the sequence data, and incorporates them into the network structure for training. First, the optimal feature representation is obtained by using the feature matrix as a trainable parameter, and CNN is used to extract features for a single variable and a single time period. Second, a segmentation method is used to extract comprehensive semantic information from the time series, and self-attention mechanisms, MLP, and gated attention mechanisms are used to extract further features. Finally, real-time prediction of RRH traffic and the number of edge users with high dynamics is achieved.

[0073] It can process multiple groups of base stations simultaneously, helping the model to discover hidden connections from these different dimensions and achieve more accurate predictions. The framework for highly dynamic traffic and user population prediction can be flexibly expanded to cope with different network scenarios. By predicting RRH traffic and edge user numbers, the network can achieve more intelligent resource allocation, improving overall network efficiency and user experience.

[0074] Since the data traffic and user connections generated in the RRH are highly dynamic at different times and places, the connection characteristics of the RRH are affected by the time context (for example, weekdays or weekends), personnel mobility, etc.; in multivariate time series prediction, the limitation of using only time series data as features is that its dimension is low, and multiple time series at adjacent time points are similar and lack distinction. This spatiotemporal similarity means that it is difficult for the prediction model to distinguish sample data at different times and spaces, so it is impossible to accurately grasp which time series each sample data comes from. The information contained in a single time point in the time series is relatively thin, and the traffic and number of connections often show mutation characteristics, that is, there are sharp increases or decreases. It is often difficult to capture this mutation feature by simply relying on information at the current time point and its surrounding time points. The present invention proposes a high-dynamic traffic and edge user number prediction method based on multivariate data fusion to solve the problems existing under these three challenges. The method can perform real-time traffic and number prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0075] Figure 1 It is a framework for predicting RRH traffic and edge user numbers. Figure 2 is the feature embedding flowchart, Figure 3 It is a prediction flow chart. DETAILED DESCRIPTION

[0076] like Figure 1-3 The method for predicting high dynamic traffic and edge user number based on multivariate data fusion of the present invention specifically includes the following steps:

[0077] Step 1: Collect time series data on the number of RRH edge users and high-dynamic traffic;

[0078] Step 2: Convert the historical RRH traffic and edge user number sequence data into embedding matrix features;

[0079] Step 3: Predict future RRH traffic and number of edge users based on the embedded matrix features.

[0080] Step 1 Specific details:

[0081] Acquires multiple heterogeneous data including RRH connection data, user attribute data, and traffic usage data, and stores them in traffic tables and user attribute tables;

[0082] The data stored in the traffic table includes the user's anonymous ID, the connection start and end time, the downlink and uplink traffic volume, and the ID of the connected RRH;

[0083] The data stored in the user attribute table includes the user's anonymous ID field, the time of the month, and the longitude and latitude information of the RRH to which the user is connected;

[0084] A new record is generated when a user connects to a new RRH or the connection duration exceeds a certain threshold;

[0085] For each hour, the traffic table and user attribute table are linked using the anonymous user ID field. Based on the RRH ID, the connection start and end times for each RRH are obtained and the connection duration is calculated. The traffic volume is calculated based on the sum of the downlink and uplink traffic volumes, along with the longitude and latitude of the connected RRH. The number of RRH-connected users is calculated by counting the number of records for a specific RRH within a time period using the anonymous user ID.

[0086] The time series tensor X is introduced to represent the characteristic information of each RRH traffic and edge user number sequence t Specifically, the sequence data is extracted at an hourly granularity, including all edge base station traffic and the number of users within an hour:

[0087]

[0088] Traffic and user represent each base station connection information extracted from the traffic table and attribute table, including the start and end time, the amount of traffic generated during the connection, and the number of users at the time of the connection. t represents the tth hour, n represents the total number of n RRHs. During this time period, traffic represents the traffic volume of an RRH, and user represents the number of edge users of an RRH.

[0089] The time order sequence tensor L is additionally introduced to express the time order feature t Specifically, for the time series tensor X t Under the premise of extracting sequence data in hourly fine granularity, considering the mean, variance and median of all edge base station traffic and number of users within an hour, the change trend sequence tensor N of the overall change trend of RRH traffic and edge user number between each time period is obtained. t ::

[0090] N t =Concat(X mean ,X sd ,X mid )

[0091] in, They represent the mean, variance and median of RRH traffic and edge user number respectively.

[0092] The number of RRH clusters is multiplied by the feature sequence feature to form the sequence data of historical RRH traffic and number of edge users, which serves as the input of the feature embedding layer:

[0093] Given a time series tensor X containing P time periods t , obtain spatiotemporal series data X t-P:t =P×X t .

[0094] Given a time sequence tensor L containing P time periods t ,, obtain time sequence data L t-P:t =P×L t .

[0095] Given a trend sequence tensor N containing P time periods t , obtain the change trend sequence data N t-P:t =P×N t .

[0096] Step 2 Specific details:

[0097] The input embedding layer is given by the formula:

[0098] E t =CNN embedding (X t-P:t ,L t-P:t ,N t-P:t )

[0099] The historical RRH traffic and edge user number sequence matrix X t-P:t ,L t-P:t and N t-P:t Convert to embedding matrix E t ∈R D ×N , where D represents the embedding dimension, CNN(·) represents the convolutional layer, and this model is implemented using a 1×1 convolutional network.

[0100] Two trainable randomly initialized matrices S and T are introduced to represent the spatial and temporal features of the time series respectively. The embedding matrix and the spatiotemporal feature expression matrix S and T are expressed by the formula

[0101] M t =Concat(E t ,S t ,T t )

[0102] Combine and generate mixed features M tIn the spatial feature matrix S, the latitude and longitude information of RRH in the time series of different RRH traffic and edge users is included; in the time feature T, the same time point between different RRH traffic and edge users time series shares the same feature value. For example, the time t of the i-th and i+1-th RRHs corresponds to and are the same.

[0103] In obtaining the feature M t After that, it passes through K convolutional layers with residual connections. Each time it passes through a layer of the network, the residual connection will superimpose the output of the previous layer of the network on the input of the next layer of the network. The following formula shows how the k-th convolutional layer is calculated:

[0104]

[0105] Finally, according to the formula:

[0106]

[0107] The output layer generates the embedding matrix Z corresponding to the RRH traffic and the number of edge users t ∈R P×N , as the input of the prediction module.

[0108] Step 3 Specific details:

[0109] Embed RRH traffic and number of people into matrix features Z t Divide into different small time segments. When deeply analyzing time series data, an effective method is to divide the input features into smaller time segments to capture the local time dynamics within the series in more detail;

[0110] Consider the input feature Z t ∈R P×N By dividing each variable into time segments with a step size of s, the input features can be converted into Z t ∈R C×L×N In the form of, where L represents the length of each time segment, it is the total number of time segments divided in a variable, which is calculated as

[0111] Z after segmentation t ∈R C×L×N Will be rearranged to Z t ∈R N×C×l .

[0112] The rearranged three-dimensional feature tensor Z t ∈R N×C×L Get the feature tensor P t ∈R N×C×D, where D represents the target dimension of the feature tensor:

[0113] The rearranged three-dimensional feature tensor Z t ∈R N×C×L , through a fully connected layer, the high-dimensional features of each sample are matched to a low-dimensional embedding vector space to obtain the feature tensor P t ∈R N×C×D , where D represents the target dimension of the feature tensor.

[0114] The feature tensor P is extracted by stacking multiple hybrid MLP layers. t ∈R N×C×D The characteristics of the RRH are used to predict the future RRH traffic and number of edge users, specifically:

[0115] A single-layer hybrid MLP architecture is used to extract the correlation features between time segments and the hidden feature correlation within a time segment respectively;

[0116] First, extract relevant features across time segments:

[0117] For the feature tensor P t Perform normalization processing.

[0118] Normalize the feature tensor P t ∈R N×C×D As the input of the self-attention mechanism.

[0119] The attention weight matrix between different time segments is obtained through the self-attention mechanism. The weight matrix is ​​multiplied by the normalized feature tensor and activated by the Softmax function. The activated result is passed through the Linear layer to obtain the output of the self-attention mechanism network.

[0120] Set the weight matrix size of the shared MLP to W M ∈R C×C The shared MLP network shares parameters between different variables and different features within the time slice, W M Learn the correlation between different time segments through linear transformation;

[0121] At the same time, the normalized embedding vector P t ∈E N×C×D Transformed into feature tensor P by matrix transposition t ∈R N×D×C , as the input of the MLP network, input feature P t is converted to R N×D×C , where C is the number of time segments, and the weight matrix acts on the last dimension of the time segment dimension, matching the input dimension;

[0122] The output of the shared MLP network is processed by the gated attention mechanism unit, and a weight matrix W is learned through the gated attention unit. G ∈R N×C×D To amplify important features and suppress irrelevant features. The calculation method of this weight matrix is ​​as follows:

[0123] W G =softmax(FC(P t ))

[0124] Where FC(·) represents the linear layer, P t Represents the output of the previous layer of network. The output of the gated attention unit is obtained by paying attention to the weight W G and the hidden tensor P output from the previous layer of the network t Dot product gives:

[0125]

[0126] Convert the output of the gated attention mechanism unit to embed P t ∈R N×D×C Transform to P by matrix transposition t ∈R N×C×D ;

[0127] The result of the conversion P t ∈R N×C×D Added to the output of the self-attention network and normalized.

[0128] The normalized output is residually connected to the input of the module. Each time a layer of the network is passed, the residual connection will superimpose the output of the previous layer of the network on the input of the next layer of the network, and finally output the relevant features across time segments.

[0129] Output the relevant features P across time segments t ∈R N×C×D .

[0130] Secondly, the extraction of hidden feature correlation within the time segment is realized.

[0131] For the correlation features P across time segments t ∈R N×C×D Perform normalization processing.

[0132] The normalized results are passed through a shared MLP unit, which learns the correlation between different embedded features within a time slice and shares parameters across different variables and time slices. The weight matrix of this network is B×B.

[0133] The output of the shared MLP network is processed by the gated attention mechanism unit and then residually connected. Each time it passes through a layer of the network, the residual connection will superimpose the output of the previous layer of the network on the input of the next layer of the network, and finally output the hidden features within the time segment.

[0134] Finally, through the multi-layer hybrid MLP layer, the output feature tensor P is obtained t ∈R N×C×D ;

[0135] Through an MLP network, the output feature tensor P t ∈R N×C×D Perform restoration to predict future RRH traffic and the number of edge users:

[0136] The output feature tensor P t ∈R N×C×D Converted into a two-dimensional form P by unfolding operation t ∈R N×(DC) The purpose of this conversion is to adjust the feature tensor into a format suitable for linear layer processing.

[0137] The expanded features are processed through a linear layer to learn the matching function from the current feature space to the prediction space. The output dimension of the linear layer is set to N×F to directly correspond to the dimension of the target prediction value.

[0138] The linear layer outputs an intermediate feature representation P t ∈R N×F .

[0139] The output of the linear layer P t ∈R N×F Transpose to obtain the final RRH traffic and edge user number prediction result Y∈R F ×N .

[0140] It can process multiple groups of base stations simultaneously, helping the model to discover hidden connections from these different dimensions and achieve more accurate predictions. The framework for highly dynamic traffic and user population prediction can be flexibly expanded to cope with different network scenarios. By predicting RRH traffic and edge user numbers, the network can achieve more intelligent resource allocation, improving overall network efficiency and user experience.

[0141] Although the present invention is presented in a large-scale cellular network scenario as an example, it is not intended to limit the present invention. Researchers and users in the field of the present invention are welcome to make changes to the structure, method, and function combination based on the embodiments of the present invention. These changes and expansions may be made as needed without departing from the spirit and scope of the present invention. Therefore, the scope of protection of the present invention shall be based on the definition of the claims.

Claims

1. A highly dynamic traffic and edge user number prediction method based on multivariate data fusion, characterized by: The following steps are involved: Step 1: Obtain time series data of historical RRH traffic and edge user numbers; Step 2: Convert the historical RRH traffic and edge user number sequence data into embedding matrix features; Step 3: Predict future RRH traffic and number of edge users based on the embedded matrix features.

2. The method for predicting highly dynamic traffic and edge user numbers based on multivariate data fusion according to claim 1, characterized in that: Step 1 specifically includes: Acquires multiple heterogeneous data including RRH connection data, user attribute data, and traffic usage data, and stores them in traffic tables and user attribute tables; Extract all RRH traffic sizes and the number of connected users in one hour from the traffic table and user attribute table in hourly fine-grained manner to obtain the time series tensor X t : Where t represents the tth hour, n represents the total number of n RRHs, traffic represents the traffic volume of an RRH during this time period, and user represents the number of edge users of an RRH. For the time series tensor X of RRH traffic and edge user number t , extracting the time order sequence tensor L that expresses the time order characteristics t ; For the time series tensor X of RRH traffic and edge user number t , on the premise of hourly fine-grained extraction, considering the mean, variance and median of all RRH traffic and edge user numbers within an hour, the change trend sequence tensor N of the overall change trend of RRH traffic and edge user numbers between each time period is obtained t : in, Represent the mean, variance and median of RRH traffic and edge user number respectively; Multiply each tensor by the number of time periods to form the sequence data of RRH traffic and edge user number: Given a time series tensor X containing P time periods t , obtain spatiotemporal series data X t-P:t =P×X t ; Given a time sequence tensor L containing P time periods t ,, obtain time sequence data L t-P:t =P×L t ; Given a trend sequence tensor N containing P time periods t , obtain the change trend sequence data N t-P:t =P×N t .

3. The method for predicting highly dynamic traffic and edge user numbers based on multivariate data fusion according to claim 1 is characterized by: Step 2 specifically includes: The embedding layer is constructed by the formula: E t =CNN embedding (X t-P:t ,L t-P:t ,N t-P:t ) The sequence data X t-P:t , L t-P:t and N t-P:t Convert to embedding matrix E t ∈R P×N ,CNN(·) represents the convolutional layer, where P represents P time periods starting from time period t, and N represents N RRHs; Two trainable randomly initialized matrices S and T are introduced to represent the time series data X t-P:t The spatial and temporal features of t ∈R P×N And the spatial feature matrix S and the temporal feature matrix T are expressed by the formula M t =Concat(E t ,S t ,T t ) Combine and generate mixed features M t The spatial feature matrix S contains the latitude and longitude information of the RRHs in the time series of different RRH traffic and edge user numbers. In the temporal feature matrix T, the same time points between the time series data of different RRH traffic and edge user numbers share the same eigenvalue. We will get the feature M t , convolution operation is performed through K convolutional layers with residual connections; According to the formula: The output layer generates the corresponding embedding matrix feature Z of RRH traffic and edge user number t ∈R P×N ; Wherein, P represents P time periods starting from time period t, and N represents N RRHs.

4. The method for predicting highly dynamic traffic and edge user numbers based on multivariate data fusion according to claim 3 is characterized by: We will get the feature M t , the convolution operation is performed through K convolutional layers with residual connections, specifically: Each time a convolutional network is passed, the residual connection will superimpose the output of the previous convolutional network to the input of the next convolutional network; the calculation method of the k-th convolutional layer is:

5. The method for predicting highly dynamic traffic and edge user numbers based on multivariate data fusion according to claim 1 is characterized by: Step 3 specifically includes: Embed the RRH traffic and edge user number into matrix features Z t Divide into different small time segments: By dividing each variable into time segments with a step size of s, the input embedding matrix feature Z t ∈R P×N Convert to Z t ∈R C×L×N In the form of, where L represents the length of each time segment, C is the number of time segments divided in a variable, and the calculation method is The divided Z t ∈R C×L×N Rearrange to Z t ∈R N×C×L ; The rearranged three-dimensional feature tensor Z t ∈R N×C×L Get the feature tensor P t ∈R N×C×D , where D represents the target dimension of the feature tensor; The feature tensor P is extracted by stacking multiple single-layer hybrid MLP layers. t ∈R N×C×D characteristics and predict future RRH traffic and edge user numbers.

6. The method for predicting highly dynamic traffic and edge user numbers based on multivariate data fusion according to claim 5, characterized in that: The rearranged three-dimensional feature tensor Z t ∈R N×C×L The characteristic tensor P t ∈R N×C×D , specifically: The rearranged three-dimensional feature tensor Z t ∈R N×C×L , through a fully connected layer, the high-dimensional feature tensor of each sample is matched to a low-dimensional feature tensor to obtain the feature tensor P t ∈R N×C×D .

7. The method for predicting highly dynamic traffic and edge user numbers based on multivariate data fusion according to claim 5 is characterized by: The feature tensor P is extracted by stacking multiple single-layer hybrid MLP layers. t ∈R N×C×D The characteristics of the RRH are used to predict the future RRH traffic and number of edge users, specifically: A single-layer hybrid MLP architecture is used to extract relevant features across time segments and hidden features within a time segment respectively; The extracted correlation features across time segments and hidden features within time segments are respectively stacked into a multi-layer hybrid MLP using the residual connection step; Through multiple layers of mixed MLP layers, the output feature tensor P is obtained t ∈R N×C×D ; Through an MLP network, the output feature tensor P t ∈R N×C×D Perform restoration to predict future RRH traffic and the number of edge users.

8. The method for predicting highly dynamic traffic and edge user numbers based on multivariate data fusion according to claim 7, characterized in that: A single-layer hybrid MLP architecture is used to extract relevant features across time segments, specifically: For the feature tensor P t Perform normalization processing; The normalized feature tensor is used as the input of the self-attention mechanism. The attention weight matrix between different time segments is obtained through the self-attention mechanism to measure the feature importance between time segments. The weight matrix is ​​multiplied by the normalized feature tensor and activated by the Softmax function. The activated result is passed through the Linear layer to obtain the output of the self-attention mechanism network. Set the weight matrix size of the shared MLP network to W M ∈R C×C , the shared MLP network shares parameters between different variables and different features within the time slice, W M Learn the correlation between different time segments through linear transformation; At the same time, the normalized feature tensor P t ∈R N×C×D Transformed into feature tensor P by matrix transposition t ∈R N×D×C , as the input of the MLP network, input feature P t is converted to R N×D×C , where C is the number of time segments, and the weight matrix acts on the last dimension of the time segment dimension, matching the input dimension; The output of the shared MLP network is processed by the gated attention mechanism unit, and a weight matrix W is learned through the gated attention unit. G ∈R N×C×D To amplify important features and suppress irrelevant features, the weight matrix is ​​calculated as follows: W G =softmax(FC(P t )) Where FC(·) represents the linear layer; The output of the gated attention unit is obtained by adding the attention weights W G and the hidden tensor P output from the previous layer of the network t Dot product gives: Convert the output of the gated attention mechanism unit to the feature tensor P t ∈R N×D×C Transform to P by matrix transposition t ∈R N×C×D ; The result of the conversion P t ∈R N×C×D Add to the output of the self-attention network and normalize; The normalized result is residually connected to the input of the self-attention mechanism. Each time a layer of the network is passed, the residual connection will superimpose the output of the previous layer of the network on the input of the next layer of the network, and finally output the relevant features across time segments.

9. The method for predicting highly dynamic traffic and edge user numbers based on multivariate data fusion according to claim 7, characterized in that: A single-layer hybrid MLP architecture is used to extract the hidden feature correlation within the time segment, specifically: For the correlation features P across time segments t ∈R N×C×D Perform normalization processing; The normalized results are passed through a shared MLP unit, which learns the correlation between different embedded features within a time slice and shares parameters across different variables and time slices. The weight matrix size of the network is B×B. The output of the shared MLP network is processed by the gated attention mechanism unit and then residually connected. Each time it passes through a layer of the network, the residual connection will superimpose the output of the previous layer of the network on the input of the next layer of the network; finally, the hidden features within the time segment are output.

10. The method for predicting highly dynamic traffic and edge user numbers based on multivariate data fusion according to claim 7, characterized in that: Through an MLP network, the output feature tensor P t ∈R N×C×D Perform restoration and predict future RRH traffic and edge user numbers, specifically: The output feature tensor P t ∈R N×C×D Converted into a two-dimensional form P by unfolding operation t ∈R N×(DC) ; After expansion, P t ∈R N×(DC) Through a linear layer processing, the matching function from the current feature space to the prediction space is learned. The output dimension of the linear layer is set to N×F to directly correspond to the dimension of the target prediction value; The linear layer outputs an intermediate feature representation P t ∈R N×F ; The output of the linear layer P t ∈R N×F Transpose to obtain the final RRH traffic and edge user number prediction result Y∈R F×N .

Images