Cellular base station network flow prediction method based on Mama and graph neural network

Abstract

The invention discloses a cellular base station network traffic prediction method based on Mama and a graph neural network, which belongs to the field of deep learning and base station traffic prediction, and comprises a linear embedded layer module, an encoder module, a double-view graph neural network module, a decoder module and an output layer module. The method specifically comprises the following steps: (1) coding base station flow data by a linear embedded layer module to obtain coding information of the base station flow data; (2) the encoder module extracts time characteristics of the base station flow data; (3) extracting spatial features of the base station flow data by a double-view image neural network module; (4) the decoder module fuses the time characteristics and the space characteristics of the base station flow data; and (5) an output layer module calculates and obtains a prediction result of the base station flow data according to the fusion result of the time features and the space features. According to the method, accurate long-term prediction of the network flow of the cellular base station is realized under the condition of fully considering the space-time correlation of the flow of the base station.

Description

Technical Field

[0001] The present invention belongs to the field of deep learning and base station traffic prediction, and in particular to a cellular base station network traffic prediction method based on Mamba and graph neural network. Background Art

[0002] With the rapid development of 5G technology, people's lives have undergone tremendous changes. According to a report from Xinhua News Agency, the number of 5G base stations in China has exceeded 3.377 million. The number of 5G users has exceeded 805 million. The explosive growth of traffic has brought huge challenges to the prediction of cellular base station network traffic.

[0003] Accurate long-term cellular traffic forecasting is critical to promoting energy efficiency optimization and reducing carbon emissions, as well as estimating crowd flow trends. For example, during periods of reduced traffic demand, operators can turn on the sleep mode of base stations, which helps achieve the communications industry's carbon peak and carbon neutrality goals. In addition, long-term cellular traffic forecasting can help managers detect rapid gatherings of people, so that they can take appropriate security measures to prevent possible public safety incidents.

[0004] Although Transformer-based models perform well in handling long-term dependencies, their quadratic computational complexity leads to low efficiency and high overhead. In addition, existing studies often fail to fully consider the spatiotemporal correlation of cellular network traffic, limiting the further improvement of prediction accuracy.

[0005] How to overcome the above common problems in the long-term prediction of base station traffic and improve the prediction accuracy has become a technical challenge that needs to be overcome urgently in the current field of base station traffic prediction technology. Summary of the invention

[0006] In order to solve the above problems, the present invention provides a cellular base station network traffic prediction method based on Mamba and graph neural network to achieve long-term and accurate prediction of base station traffic.

[0007] To achieve the above-mentioned object, the present invention provides a cellular base station network traffic prediction method based on Mamba and graph neural network, comprising a linear embedding layer module, an encoder module, a dual-view graph neural network module, a decoder module and an output layer module;

[0008] The specific steps include:

[0009] Step S1: The linear embedding layer module encodes the base station traffic data through a linear layer and maps the base station traffic data into a high-dimensional vector;

[0010] Step S2: input the high-dimensional vector obtained in step S1 into the encoder module, and capture the time correlation between the base station traffic data through the bidirectional Mamba in the encoder module;

[0011] Step S3: Input the temporal correlation result obtained in step S2 into the dual-view graph neural network module, capture the spatial correlation of traffic data between adjacent base stations through the graph convolution network in the dual-view graph neural network module, and obtain the spatial correlation of traffic data between distant base stations through the graph attention network;

[0012] Step S4: input the spatial correlation result obtained in step S3 into the decoder module, and fuse the spatial correlation and time correlation between the base station traffic data through bidirectional Mamba;

[0013] Step S5: The temporal and spatial fusion features of the base station traffic obtained in step S4 are input into the output layer module to obtain the prediction result.

[0014] Preferably, in step S1, the base station traffic data X∈R is transformed into B×L×N Mapping to high-dimensional space X emb ∈R B×N×D , the formula is as follows:

[0015] X emb =Linear(X);

[0016] Among them, X is the base station traffic data, B is the batch size, N is the number of base stations, L is the historical traffic length, and D is the feature dimension after mapping through the linear layer;

[0017] After the linear layer, the feature dimension of the base station traffic data changes from L to D.

[0018] Preferably, in step S2, the following steps are specifically included:

[0019] S21: The encoder module converts the output of the linear layer X emb ∈R B×N×D , the temporal correlation between the base station traffic data is captured by bidirectional Mamba, and H encmamba ∈R B×N×D , expressed as:

[0020] X encfilp =Filp(X emb );

[0021] H encmamba =Mamba(X emb )+Mamba(X encfilp );

[0022] Filp() means flipping, Mamba() means extracting features using Mamba, and X encfilp Represents X emb The value after flipping;

[0023] S22: Encoder module will bidirectional Mamba output H encmamba ∈R B×N×D After residual connection, the output H is obtained encres ∈R B×N×D :

[0024] H encres =H encmamba +X emb ;

[0025] S23: The encoder module connects the residual output H encres ∈R B×N×D After layer normalization and forward feedback neural network, the output H is obtained enc ∈R B×N×D :

[0026] After normalization, we get H encres '∈R B×N×D :

[0027] H encres '=LayerNorm(H encres );

[0028] The normalized H encres '∈R B×N×D After being processed by the forward feedback neural network, H encres ”∈R B×N×D :

[0029] H encres =FeedForward(H encres ');

[0030] H enc =LayerNorm(H encres ”);

[0031] Among them, LayerNorm() represents layer normalization, and FeedForward() represents the forward feedback neural network.

[0032] Preferably, in step S3, the following steps are specifically included:

[0033] S31: The graph convolutional network constructs the distance matrix A according to the following formula gcn :

[0034]

[0035] Among them, vi ,v j represents two different nodes i and j, dis(v i ,v j ) is the distance between nodes i and j, σ is the design parameter, and κ is the threshold;

[0036] Adjacency matrix A of graph attention network gat It is expressed as:

[0037]

[0038] Among them, cov() represents the covariance operator of the rank variable, σ represents the standard deviation of the rank variable, and Represents the cellular data of nodes i and j;

[0039] S32: The dual-view graph neural network module converts the encoder output H enc ∈R B×N×D The spatial correlation between base station traffic data is captured by graph convolutional network and graph attention network, and the graph convolutional network output H is obtained. gcn ∈R B×N×D And the graph attention network output H gat ∈R B×N×D :

[0040] H gcn =GCN(H,A gcn );

[0041] H gat =GAT(H,A gat );

[0042] Among them, GCN() represents graph convolutional network, GAT() represents graph attention network, B represents batch size, N represents the number of base stations, and D represents the feature dimension after mapping through the linear layer;

[0043] S33: The encoder output H enc ∈R B×N×D , graph convolutional network output H gcn ∈R B×N×D And the graph attention network output H gat ∈R B×N×D Splicing to get H∈R B×N×3D :

[0044] H = Concat(H enc ,H gcn ,H gat );

[0045] Among them, Concat() means concatenation;

[0046] S34: The concatenated output H∈R B×N×3D Use layer normalization and linear layer to initially fuse to get the output H st ∈R B ×N×3D :

[0047] The concatenated output H∈R B×N×3D After normalization, we get H'∈R B×N×3D :

[0048] H' = LayerNorm(H);

[0049] The normalized H'∈R B×N×3D Perform linear layer fusion to obtain H st ∈R B×N×3D :

[0050] H st =Linear(H');

[0051] Among them, LayerNorm() represents layer normalization, and Linear() represents the linear layer.

[0052] Preferably, in step S4, the following steps are specifically included:

[0053] S41: Output H after fusion and splicing st ∈R B×N×3D , the temporal correlation between the base station traffic data is captured by bidirectional Mamba, and H decmamba ∈R B×N×3D Specifically:

[0054] H decfilp =Filp(H st );

[0055] H decmamba =Mamba(H st )+Mamba(H decfilp );

[0056] Filp() means flipping, Mamba() means extracting features using Mamba, and H decfilp Indicates H st The value after flipping;

[0057] S42: Bidirectional Mamba output H decmamba ∈R B×N×3D After residual connection, the output H is obtained encres ∈R B×N×D :

[0058] H decres =H decmamba +Hst ;

[0059] S43: The output H of the residual connection decres ∈R B×N×3D , after layer normalization and forward feedback neural network, the output H is obtained dec ∈R B×N×3D :

[0060] H decres ∈R B×N×3D Normalize to get H decres ':

[0061] H decres '=LayerNorm(H decres );

[0062] H decres 'After the forward feedback neural network processing, H decres ”:

[0063] H decres =FeedForward(H decres ');

[0064] H dec =LayerNorm(H decres ”);

[0065] Among them, LayerNorm() represents layer normalization, and FeedForward() represents the forward feedback neural network.

[0066] Preferably, in step S5, the output layer module maps the temporal and spatial fusion features of the base station traffic into a predicted output length, specifically comprising the following steps: dec ∈R B×N×3D Transformed into prediction result Y∈R through linear layer B×T×N , specifically:

[0067] Y=Linear(H dec );

[0068] Among them, Linear() represents the linear layer, and T represents the predicted length.

[0069] Therefore, the present invention adopts the above-mentioned cellular base station network traffic prediction method based on Mamba and graph neural network, which has the following beneficial effects:

[0070] (1) Aiming at the base station scenario, the present invention uses Mamba and graph neural network models to extract more effective spatiotemporal feature information.

[0071] (2) While focusing on the temporal characteristics of base station traffic data, the present invention also considers the spatial heterogeneity of base station traffic data.

[0072] (3) The experimental results of the present invention on a real base station traffic data set show that compared with the latest prediction model, this scheme has achieved improvement in prediction accuracy.

[0073] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0074] Figure 1 It is a schematic diagram of the structure of a cellular base station network traffic prediction method based on Mamba and graph neural network of the present invention;

[0075] Figure 2 This is an architecture diagram of a cellular base station network traffic prediction method based on Mamba and graph neural network of the present invention;

[0076] Figure 3 It is a performance comparison bar chart of the present solution and other methods when the prediction length is 96 in the comparative example of the present invention;

[0077] Figure 4 It is a performance comparison bar chart of the present solution and other methods when the prediction length is 192 in the comparative example of the present invention;

[0078] Figure 5 It is a performance comparison bar chart of the present solution and other methods when the prediction length is 336 in the comparative example of the present invention;

[0079] Figure 6 It is a bar chart comparing the performance of this solution and other methods when the prediction length is 720 in the comparative example of the present invention. DETAILED DESCRIPTION

[0080] The technical solution of the present invention is further described below through the accompanying drawings and embodiments.

[0081] Unless otherwise defined, technical or scientific terms used in the present invention shall have the common meanings understood by one having ordinary skills in the field to which the present invention belongs.

[0082] The words "include" or "comprises" and the like used in the present invention mean that the elements before the word include the elements listed after the word, and do not exclude the possibility of also including other elements. The orientation or position relationship indicated by the terms "inside", "outside", "upper", "lower", etc. is based on the orientation or position relationship shown in the drawings, which is only for the convenience of describing the present invention and simplifying the description, and does not indicate or imply that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation. Therefore, it cannot be understood as a limitation of the present invention. When the absolute position of the described object changes, the relative position relationship may also change accordingly. In the present invention, unless otherwise clearly specified and limited, the terms "attachment" and the like should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral body; it can be directly connected, or indirectly connected through an intermediate medium, and it can be the internal connection of two elements or the interaction relationship between two elements. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0083] Example

[0084] like Figure 1-Figure 2 As shown, a cellular base station network traffic prediction method based on Mamba and graph neural network includes a linear embedding layer module, an encoder module, a dual-view graph neural network module, a decoder module and an output layer module;

[0085] The specific steps include:

[0086] Step S1: The linear embedding layer module encodes the base station traffic data through the linear layer and maps the base station traffic data into a high-dimensional vector; in step S1, it specifically includes using the linear layer to encode the base station traffic data X∈R B×L×N Mapping to high-dimensional space X emb ∈R B×N×D , the formula is as follows:

[0087] X emb =Linear(X);

[0088] Among them, X is the base station traffic data, B is the batch size, N is the number of base stations, L is the historical traffic length, and D is the feature dimension after mapping through the linear layer; B is 32, N is 900, L is 96, and D is 512;

[0089] After the linear layer, the feature dimension of the base station traffic data changes from L to D.

[0090] Step S2: input the high-dimensional vector obtained in step S1 into the encoder module, and capture the time correlation between the base station traffic data through the bidirectional Mamba in the encoder module; in step S2, the following steps are specifically included:

[0091] S21: The encoder module converts the output of the linear layer X emb ∈R B×N×D , the temporal correlation between the base station traffic data is captured by bidirectional Mamba, and H encmamba ∈R B×N×D , expressed as:

[0092] X encfilp =Filp(X emb );

[0093] H encmamba =Mamba(X emb )+Mamba(X encfilp );

[0094] Filp() means flipping, Mamba() means extracting features using Mamba, and X encfilp Represents X emb The value after flipping;

[0095] S22: Encoder module will bidirectional Mamba output H encmamba ∈R B×N×D After residual connection, the output H is obtained encres ∈R B×N×D :

[0096] H encres =H encmamba +X emb ;

[0097] S23: The encoder module connects the residual output H encres ∈R B×N×D After layer normalization and forward feedback neural network, the output H is obtained enc ∈R B×N×D :

[0098] After normalization, we get H encres '∈R B×N×D :

[0099] H encres '=LayerNorm(H encres );

[0100] The normalized H encres '∈R B×N×D After being processed by the forward feedback neural network, H encres ”∈R B×N×D :

[0101] H encres = FeedForward(H encres ');

[0102] H enc =LayerNorm(H encres ”);

[0103] Among them, LayerNorm() represents layer normalization, and FeedForward() represents the forward feedback neural network.

[0104] Step S3: input the temporal correlation obtained in step S2 into the dual-view graph neural network module, capture the spatial correlation of traffic data between adjacent base stations through the graph convolution network in the dual-view graph neural network module, and obtain the spatial correlation of traffic data between distant base stations through the graph attention network;

[0105] In step S3, the following steps are specifically included:

[0106] S31: The graph convolutional network constructs the distance matrix A according to the following formula gcn :

[0107]

[0108] Among them, v i ,v j represents two different nodes i and j, dis(v i ,v j ) is the distance between nodes i and j, σ is the design parameter, and κ is the threshold;

[0109] Adjacency matrix A of graph attention network gat It is expressed as:

[0110]

[0111] Among them, cov() represents the covariance operator of the rank variable, σ represents the standard deviation of the rank variable, and Represents the cellular data of nodes i and j;

[0112] S32: The dual-view graph neural network module converts the encoder output H enc ∈R B×N×D The spatial correlation between base station traffic data is captured by graph convolutional network and graph attention network, and the graph convolutional network output H is obtained. gcn ∈R B×N×D And the graph attention network output H gat ∈R B×N×D :

[0113] H gcn =GCN(H,A gcn );

[0114] H gat =GAT(H,Agat );

[0115] Among them, GCN() represents graph convolutional network, GAT() represents graph attention network, B represents batch size, N represents the number of base stations, and D represents the feature dimension after mapping through the linear layer;

[0116] S33: The encoder output H enc ∈R B×N×D , graph convolutional network output H gcn ∈R B×N×D And the graph attention network output H gat ∈R B×N×D Splicing to get H∈R B×N×3D :

[0117] H = Concat(H enc ,H gcn ,H gat );

[0118] Among them, Concat() means concatenation;

[0119] S34: The concatenated output H∈R B×N×3D Use layer normalization and linear layer to initially fuse to get the output H st ∈R B ×N×3D :

[0120] The concatenated output H∈R B×N×3D After normalization, we get H'∈R B×N×3D :

[0121] H' = LayerNorm(H);

[0122] The normalized H'∈R B×N×3D Perform linear layer fusion to obtain H st ∈R B×N×3D :

[0123] H st =Linear(H');

[0124] Among them, LayerNorm() represents layer normalization, and Linear() represents the linear layer.

[0125] Step S4: input the spatial correlation result obtained in step S3 into the decoder module, and fuse the spatial correlation and time correlation between the base station traffic data through bidirectional Mamba;

[0126] In step S4, the following steps are specifically included:

[0127] S41: Output H after fusion and splicing st∈R B×N×3D , the temporal correlation between the base station traffic data is captured by bidirectional Mamba, and H decmamba ∈R B×N×3D Specifically:

[0128] H decfilp =Filp(H st );

[0129] H decmamba =Mamba(H st )+Mamba(H decfilp );

[0130] Filp() means flipping, Mamba() means extracting features using Mamba, and H decfilp Indicates H st The value after flipping;

[0131] S42: Bidirectional Mamba output H decmamba ∈R B×N×3D After residual connection, the output H is obtained encres ∈R B×N×D :

[0132] H decres =H decmamba +H st ;

[0133] S43: The output H of the residual connection decres ∈R B×N×3D , after layer normalization and forward feedback neural network, the output H is obtained dec ∈R B×N×3D :

[0134] H decres ∈R B×N×3D Normalize to get H decres ':

[0135] H decres '=LayerNorm(H decres );

[0136] H decres 'After the forward feedback neural network processing, H decres ”:

[0137] H decres =FeedForward(H decres ');

[0138] H dec =LayerNorm(H decres ”);

[0139] Among them, LayerNorm() represents layer normalization, and FeedForward() represents the forward feedback neural network.

[0140] Step S5: input the temporal and spatial fusion features of the base station traffic obtained in step S4 into the output layer module to obtain the prediction result;

[0141] In step S5, the output layer module maps the temporal and spatial fusion features of the base station traffic into the predicted output length, which specifically includes the following steps: dec ∈R B×N×3D Transformed into prediction result Y∈R through linear layer B ×T×N , specifically:

[0142] Y=Linear(H dec );

[0143] Among them, Linear() represents the linear layer, and T represents the predicted length.

[0144] Use evaluation indicators MAE and MSE to judge prediction accuracy;

[0145]

[0146] In the above formula, N represents the number of base stations; X i represents the base station traffic data of the i-th base station in the base station traffic data X, Y i Represents the prediction result of the i-th base station in the prediction result Y.

[0147] Comparative Example

[0148] After training, the method was verified on a real data set, which contains 900 base stations. The prediction window lengths were set to 96, 192, 336, and 720 respectively. The comparison of the prediction results of the base station traffic data input to the base station by this scheme and Itransformer, PatchTST, S-mamba, DLinear, and TimeMachine models under the above prediction window lengths is shown in the figure below. Figure 3 , Figure 4 , Figure 5 , Figure 6 shown.

[0149] Therefore, the present invention adopts the above-mentioned cellular base station network traffic prediction method based on Mamba and graph neural network, which realizes accurate long-term prediction of cellular base station network traffic while fully considering the temporal and spatial correlation of base station traffic.

[0150] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that they can still modify or replace the technical solution of the present invention with equivalents, and these modifications or equivalent replacements cannot cause the modified technical solution to deviate from the spirit and scope of the technical solution of the present invention.

Claims

1. A cellular base station network traffic prediction method based on Mamba and graph neural network, characterized by: It includes a linear embedding layer module, an encoder module, a dual-view graph neural network module, a decoder module and an output layer module; The specific steps include: Step S1: The linear embedding layer module encodes the base station traffic data through a linear layer and maps the base station traffic data into a high-dimensional vector; Step S2: input the high-dimensional vector obtained in step S1 into the encoder module, and capture the time correlation between the base station traffic data through the bidirectional Mamba in the encoder module; Step S3: Input the temporal correlation result obtained in step S2 into the dual-view graph neural network module, capture the spatial correlation of traffic data between adjacent base stations through the graph convolution network in the dual-view graph neural network module, and obtain the spatial correlation of traffic data between distant base stations through the graph attention network; Step S4: input the spatial correlation result obtained in step S3 into the decoder module, and fuse the spatial correlation and time correlation between the base station traffic data through bidirectional Mamba; Step S5: The temporal and spatial fusion features of the base station traffic obtained in step S4 are input into the output layer module to obtain the prediction result.

2. A method for predicting network traffic of cellular base stations based on Mamba and graph neural network according to claim 1, characterized in that: In step S1, the base station traffic data X∈R is transformed into B×L×N Mapping to high-dimensional space X emb ∈R B×N×D , the formula is as follows: X emb =Linear(X); Among them, X is the base station traffic data, B is the batch size, N is the number of base stations, L is the historical traffic length, and D is the feature dimension after mapping through the linear layer; After the linear layer, the feature dimension of the base station traffic data changes from L to D.

3. A method for predicting cellular base station network traffic based on Mamba and graph neural network according to claim 1, characterized in that: In step S2, the following steps are specifically included: S21: The encoder module converts the output of the linear layer X emb ∈R B×N×D , the temporal correlation between base station traffic data is captured by bidirectional Mamba and feedforward neural network, and H encmamba ∈R B×N×D , expressed as: X encfilp =Filp(X emb ); H encmamba =Mamba(X emb )+Mamba(X encfilp ); Filp() means flipping, Mamba() means extracting features using Mamba, and X encfilp Represents X emb The value after flipping; S22: Encoder module will bidirectional Mamba output H encmamba ∈R B×N×D After residual connection, the output H is obtained encres ∈R B ×N×D : H encres =H encmamba +X emb ; S23: The encoder module connects the residual output H encres ∈R B×N×D After layer normalization and forward feedback neural network, the output H is obtained enc ∈R B×N×D : After normalization, we get H encres '∈R B×N×D : H encres '=LayerNorm(H encres ); The normalized H encres '∈R B×N×D After being processed by the forward feedback neural network, H encres ”∈R B×N×D : H encres ”=FeedForward(H encres '); H enc =LayerNorm(H encres ”); Among them, LayerNorm() represents layer normalization, and FeedForward() represents the forward feedback neural network.

4. A method for predicting network traffic of cellular base stations based on Mamba and graph neural network according to claim 3, characterized in that: In step S3, the following steps are specifically included: S31: The graph convolutional network constructs the distance matrix A according to the following formula gcn : Among them, v i ,v j represents two different nodes i and j, dis(v i ,v j ) is the distance between nodes i and j, σ is the design parameter, and κ is the threshold; Adjacency matrix A of graph attention network gat It is expressed as: Among them, cov() represents the covariance operator of the rank variable, σ represents the standard deviation of the rank variable, and Represents the cellular data of nodes i and j; S32: The dual-view graph neural network module converts the encoder output H enc ∈R B×N×D The spatial correlation between base station traffic data is captured by graph convolutional network and graph attention network, and the graph convolutional network output H is obtained. gcn ∈R B×N×D And the graph attention network output H gat ∈R B×N×D : H gcn =GCN(H,A gcn ); H gat =GAT(H,A gat ); Among them, GCN() represents graph convolutional network, GAT() represents graph attention network, B represents batch size, N represents the number of base stations, and D represents the feature dimension after mapping through the linear layer; S33: The encoder output H enc ∈R B×N×D , graph convolutional network output H gcn ∈R B×N×D And the graph attention network output H gat ∈R B×N×D Splicing to get H∈R B×N×3D : H=Concat(H enc ,H gcn ,H gat ); Among them, Concat() means concatenation; S34: The concatenated output H∈R B×N×3D Use layer normalization and linear layer to initially fuse to get the output H st ∈R B×N×3D : The concatenated output H∈R B×N×3D After normalization, we get H'∈R B×N×3D : H' = LayerNorm(H); The normalized H' is fused with linear layers to obtain H st ∈R B×N×3D : H st =Linear(H'); Among them, LayerNorm() represents layer normalization, and Linear() represents the linear layer.

5. A method for predicting cellular base station network traffic based on Mamba and graph neural network according to claim 4, characterized in that: In step S4, the following steps are specifically included: S41: Output H after fusion and splicing st ∈R B×N×3D , the temporal correlation between the base station traffic data is captured by bidirectional Mamba, and H decmamba ∈R B×N×3D Specifically: H decfilp =Filp(H st ); H decmamba =Mamba(H st )+Mamba(H decfilp ); Filp() means flipping, Mamba() means extracting features using Mamba, and H decfilp Indicates H st The value after flipping; S42: Bidirectional Mamba output H decmamba ∈R B×N×3D The converged output obtained through the residual connection is: H decres =H decmamba +H st ; S43: The output H of the residual connection decres ∈R B×N×3D , after layer normalization and forward feedback neural network, the output H is obtained dec ∈R B×N×3D : H decres ∈R B×N×3D Normalize to get H decres ': H decres '=LayerNorm(H decres ); H decres 'After the forward feedback neural network processing, H decres ”: H decres ”=FeedForward(H decres '); H dec =LayerNorm(H decres ”); Among them, LayerNorm() represents layer normalization, and FeedForward() represents the forward feedback neural network.

6. A method for predicting cellular base station network traffic based on Mamba and graph neural network according to claim 5, characterized in that: In step S5, the output layer module maps the temporal and spatial fusion features of the base station traffic into the predicted output length, which specifically includes the following steps: dec ∈R B×N×3D Transformed into prediction result Y∈R through linear layer B ×T×N , specifically: Y=Linear(H dec ); Among them, Linear() represents the linear layer, and T represents the predicted length.

Images