A multi-step traffic prediction method for cellular networks

Abstract

This invention discloses a multi-step traffic prediction method for cellular networks, specifically including the following steps: obtaining historical cellular network traffic sequences from a server and preprocessing them to obtain historical traffic input sequences; inputting the historical traffic input sequences into a global spatial Transformer after hybrid spatiotemporal embedding, and using a multi-head self-attention mechanism to obtain a spatial prediction representation; dividing the historical traffic input sequences into multiple time segments and inputting them into a segmented cyclic encoding and decoding structure to obtain a temporal prediction representation; fusing the spatial and temporal prediction representations to obtain the final traffic prediction result. This invention enhances variable representation capabilities through a hybrid spatiotemporal embedding mechanism, captures global spatial dependencies using a global spatial Transformer, and mines long-term temporal dynamic features using a segmented cyclic structure; it introduces a gating fusion mechanism to achieve adaptive fusion of spatial and temporal prediction representations, thereby improving the representation capability of complex spatiotemporal patterns and increasing the accuracy of multi-step traffic prediction.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a multi-step traffic prediction method for cellular networks. Background Technology

[0002] With the widespread deployment of fifth-generation and next-generation mobile communication systems, the rapid development of services such as the Internet of Things and ultra-high-definition video, and the continuous evolution of new communication technologies such as smart reflectors, cellular network traffic exhibits massive, dynamic, and heterogeneous characteristics in both time and space. Accurate cellular network traffic forecasting, as a crucial supporting technology for network self-organization and intelligent management, is of great significance for operators to implement forward-looking resource scheduling, dynamic base station hibernation, and refined capacity planning. Compared to single-step forecasting that only predicts traffic demand at the next moment, multi-step traffic forecasting can provide future traffic change trends within a longer decision window, serving as a vital support for achieving intelligent network operation and maintenance.

[0003] Early cellular network traffic prediction primarily relied on traditional statistical or classical machine learning methods such as the Autoregressive Integrated Moving Average (ARIMA) model and Support Vector Regression (SVR). However, these methods are typically based on stationarity assumptions or artificial feature engineering, making it difficult to effectively capture the strong nonlinearity, spatiotemporal tidal variations, and complex abrupt changes behind cellular traffic.

[0004] In recent years, deep learning has made significant progress in the field of spatiotemporal sequence prediction. The combination of Graph Convolutional Networks (GCNs) with Recurrent Neural Networks (RNNs) or Convolutional Neural Networks (CNNs) has become a common architecture in cellular network traffic prediction. These methods capture spatial correlations between different regions or base stations through spatial topology graphs and utilize temporal networks to capture the dependencies of traffic changes over time. Furthermore, with the development of the Transformer architecture in temporal modeling, models based on self-attention mechanisms, with their ability to model long-range dependencies, are also increasingly being applied to cellular network traffic prediction tasks.

[0005] However, existing methods still face the following two core challenges when dealing with long-term, multi-step traffic prediction tasks in cellular networks:

[0006] 1. Insufficient modeling of spatiotemporal heterogeneity. Cellular traffic is not only affected by long-term spatial non-stationarity caused by geographical location (such as commercial areas and residential areas), but also strongly constrained by temporal periodic patterns (temporal phase characteristics) such as day-night alternation and weekdays versus weekends. This interweaving of spatial attributes and temporal periods creates complex spatiotemporal heterogeneity. Existing deep spatiotemporal prediction methods often directly integrate spatial and temporal representations, or rely on fixed preset topological maps in space, making it difficult to fully characterize the dynamic evolution of traffic in different regions under different temporal phases, thus limiting prediction accuracy.

[0007] 2. Multi-step prediction suffers from error accumulation and long-term accuracy degradation. Many existing spatiotemporal prediction methods, when performing multi-step prediction, essentially still employ autoregressive recursive prediction or classic autoregressive decoding mechanisms from sequence to sequence, using the prediction result of the previous step as part of the input for the next step. Under this mechanism, small prediction errors in the early stages gradually propagate and accumulate as the prediction step size increases, leading to a decrease in accuracy in the later stages of prediction. Summary of the Invention

[0008] To address the technical problems of low accuracy and insufficient long-term prediction stability in multi-step traffic prediction for cellular networks in existing technologies, this invention proposes a multi-step traffic prediction method for cellular networks. This method enhances variable representation capabilities through a hybrid spatiotemporal embedding mechanism, captures global spatial dependencies using a global spatial Transformer, and mines long-term temporal dynamic features using a piecewise cyclic structure. Simultaneously, a gating fusion mechanism is introduced to achieve adaptive fusion of spatial and temporal prediction representations, thereby improving the ability to represent complex spatiotemporal patterns and enhancing the accuracy and long-term prediction stability of multi-step traffic prediction.

[0009] To achieve the above objectives, the present invention provides the following technical solution:

[0010] A multi-step traffic prediction method for cellular networks specifically includes the following steps:

[0011] S1: Obtain historical cellular network traffic sequences from the server and preprocess them to obtain historical traffic input sequences;

[0012] S2: After the historical traffic input sequence is mixed with spatiotemporal embedding, it is input into the global spatial Transformer, and spatial prediction representation is obtained by using a multi-head self-attention mechanism;

[0013] S3: Divide the historical traffic input sequence into multiple time segments and input them into a segmented cyclic encoding and decoding structure to obtain a time prediction representation;

[0014] S4: The final traffic prediction result is obtained by fusing the spatial prediction representation and the temporal prediction representation.

[0015] In summary, by adopting the above technical solution, the present invention has at least the following beneficial effects compared with the prior art:

[0016] (1) A hybrid spatiotemporal embedding mechanism was designed. Unlike traditional methods that mainly organize the embedding of multiple spatial region flow sequences according to time steps, this invention adopts an inverted embedding strategy, mapping the complete historical flow sequence of each spatial channel to an independent variable token, and on this basis, integrating spatial channel embedding, temporal phase embedding, and joint spatial channel-temporal phase embedding. This mechanism enhances the model's ability to represent spatiotemporal heterogeneity and nonlinear periodic characteristics while maintaining the semantic independence of flow in different spatial regions.

[0017] (2) Introduce a global spatial Transformer. By using a multi-head self-attention mechanism to model spatial correlation, without the need to pre-set a fixed spatial topology adjacency matrix, it can dynamically and adaptively capture cross-regional long-distance spatial dependencies between different spatial grids within the urban area, thereby improving the ability to model global spatial correlation patterns.

[0018] (3) Construct a segmented cyclic parallel decoding mechanism. By dividing the historical flow sequence and the future prediction interval into multiple time segments, the segmented cyclic structure is used to reduce the number of loop iterations in the time domain modeling process. In the decoding stage, a parallel multi-step forecasting (PMF) mechanism is adopted to generate the prediction results of multiple future time segments in one decoding process, avoiding the repeated use of the previous step prediction value as the subsequent input, thereby reducing the problem of error gradually propagating and accumulating as the prediction step length increases, alleviating the accuracy decay phenomenon under long prediction step length, and improving the accuracy of multi-step flow prediction.

[0019] (4) Introduce a gating fusion mechanism. By using learnable gating weights to adaptively fuse spatial and temporal prediction representations, the model can automatically adjust the contribution of spatial and temporal information in the final prediction result based on the training data, thereby enhancing the model's comprehensive representation ability of complex spatiotemporal flow patterns and improving the accuracy and stability of multi-step flow prediction. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of a multi-step traffic prediction method for cellular networks according to an exemplary embodiment of the present invention.

[0021] Figure 2 This is a schematic diagram illustrating the process of constructing historical traffic input sequences according to an exemplary embodiment of the present invention.

[0022] Figure 3 This is a schematic diagram showing the comparison results of the MAE performance curves of the multi-step flow prediction model according to an exemplary embodiment of the present invention.

[0023] Figure 4This is a schematic diagram illustrating the RMSE ablation comparison results of the Milan cellular network traffic dataset according to an exemplary embodiment of the present invention.

[0024] Figure 5 This is a schematic diagram of MAE ablation comparison results for a Milan cellular network traffic dataset according to an exemplary embodiment of the present invention.

[0025] Figure 6 This is a schematic diagram illustrating the comparison of RMSE ablation results for Beijing and Nanjing cellular network traffic datasets according to an exemplary embodiment of the present invention.

[0026] Figure 7 This is a schematic diagram comparing prediction results for 12 consecutive prediction steps according to an exemplary embodiment of the present invention.

[0027] Figure 8 This is a visual comparison diagram of the spatial distribution and absolute error of the prediction results in step 12 according to an exemplary embodiment of the present invention. Detailed Implementation

[0028] The present invention will be further described in detail below with reference to embodiments and specific implementation methods. However, this should not be construed as limiting the scope of the above-described subject matter of the present invention to the following embodiments; all technologies implemented based on the content of the present invention fall within the scope of the present invention.

[0029] In the description of this invention, it should be understood that the terms "longitudinal", "lateral", "up", "down", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0030] In existing technologies, the following prediction methods are generally used for cellular network traffic prediction:

[0031] (1) Prediction methods based on statistics and traditional machine learning

[0032] Cellular network traffic forecasting aims to predict network service load over a future period using historical communication data, serving as a crucial foundation for optimized network resource allocation, capacity planning, and intelligent operation and maintenance. Due to the combined influence of user behavior, service type, and geographic spatial distribution, cellular network traffic typically exhibits significant spatiotemporal correlation, periodicity, and non-stationarity, making the forecasting task highly complex.

[0033] Early research primarily employed statistical methods for modeling, such as autoregressive models (AR), autoregressive moving average models (ARMA), and autoregressive integral moving average models (ARIMA). These methods offer good interpretability but typically rely on stationarity assumptions, making it difficult to effectively characterize complex nonlinear variations. With the development of machine learning, methods such as support vector machines (SVM) and random forests (RF) have gradually been applied to traffic prediction tasks. However, these methods often rely on manual feature engineering, making it difficult to fully uncover the deep spatiotemporal correlations within traffic data.

[0034] (2) Prediction method based on recurrent neural network

[0035] Recurrent Neural Networks (RNNs) are a typical sequence modeling method that uses recursive propagation of hidden states over time to remember and update historical information. However, traditional RNNs are prone to gradient vanishing and gradient exploding problems during long sequence learning, thus affecting their ability to model long-term dependencies.

[0036] To address the aforementioned issues, Long Short-Term Memory (LSTM) networks were proposed and have been widely applied in time series forecasting tasks. Furthermore, some methods have achieved good results in spatiotemporal forecasting by introducing convolutional operations to jointly model spatial and temporal features.

[0037] In recent years, researchers have proposed a series of improved recurrent structures for long-term time series forecasting. Among them, the Segmented Recurrent Neural Network (SegRNN) reduces the difficulty of modeling long series to some extent by dividing them into multiple consecutive segments for modeling, while preserving long-term dependency information. Compared with traditional recursive forecasting methods, SegRNN exhibits better stability and predictive ability in long-term forecasting tasks. However, recurrent networks still have shortcomings in modeling complex spatial relationships and struggle to fully capture global spatial dependencies in cellular network traffic.

[0038] (3) Transformer-based prediction methods

[0039] The core idea of ​​Transformer is to establish global dependencies within a sequence using a self-attention mechanism. Compared to traditional recurrent structures, Transformer can directly model the relationships between arbitrary positions, demonstrating better performance in long sequence prediction tasks.

[0040] In recent years, numerous studies have introduced Transformers into the field of time series prediction. For example, probabilistic sparse attention mechanisms have been used to reduce computational complexity and improve the efficiency of long-sequence prediction; sequence decomposition and autocorrelation mechanisms have been introduced to enhance long-term dependency modeling capabilities; frequency domain decomposition has been used to further improve prediction performance; patch partitioning strategies have been employed to enhance local time series modeling capabilities; and multi-scale hybrid mechanisms have been used to achieve interactive learning of features at different time scales.

[0041] For cellular network traffic prediction tasks, researchers have further extended the Transformer to spatiotemporal modeling scenarios. For example, they used spatial attention and temporal attention to model spatial associations and temporal dependencies, respectively; and constructed a global-local spatiotemporal Transformer network to achieve joint modeling of city-level cellular network traffic. These methods can capture global spatial dependencies between distant regions to some extent, but they still have certain limitations in modeling local temporal patterns and error propagation in long-term predictions.

[0042] In this embodiment, the goal of multi-step traffic prediction for cellular networks is to construct a mapping function based on known historical traffic observation sequences. This is used to directly and in parallel predict the full-grid flow forecast results for the next k consecutive time steps. :

[0043]

[0044]

[0045] in, Indicates the future A full-grid flow prediction matrix for each time step; Represents the historical flow input sequence; Indicates the first The predicted flow vector at each time step ; Indicates the total number of grid cells; superscript This indicates the transpose operation.

[0046] like Figure 1 As shown, this invention provides a multi-step traffic prediction method for cellular networks, specifically including the following steps:

[0047] S1: Obtain historical cellular network traffic sequences from the server and preprocess them to obtain historical traffic input sequences.

[0048] In this embodiment, as Figure 2 As shown, S1 specifically includes the following steps:

[0049] S1-1: Obtain cellular network traffic data for the target area within a continuous historical time period from the server, and divide the target area into... The system uses regular grid cells and unfolds the two-dimensional flow graph of each time step into a one-dimensional flow vector according to a fixed grid order.

[0050] In this embodiment, the present invention divides the urban area according to a geospatial grid, assuming that the entire observation area is divided into... A grid of regular cells, each representing a spatial channel (or spatial node), where V represents the number of rows and H represents the number of columns. In mobile communication systems, the cellular network traffic generated by each cell within a specific time period can be measured using continuous numerical values.

[0051] set up Indicates at time step At that time, all within the entire observation area Cellular network traffic observation vectors for each grid cell. Indicates the first Each grid cell at time step Flow observations .

[0052] S1-2: Select continuous One-dimensional flow vectors at each historical time step are stacked in chronological order to obtain the historical flow input sequence.

[0053] Given continuous Historical flow observation data at each time step are used to construct a historical flow input sequence:

[0054]

[0055] in, Represents the historical flow input sequence; Indicates time step Flow observation vector; Indicates time step Flow observation vector; Indicates time step Flow observation vector; superscript This indicates the transpose operation.

[0056] S2: After the historical traffic input sequence is mixed and spatiotemporally embedded, it is input into the existing global spatial Transformer. The multi-head self-attention mechanism is used to capture the global spatial dependencies between different grid cells to obtain the spatial prediction representation.

[0057] S2-1: The inverted embedding method is used to map the complete historical traffic sequence of each grid cell in the historical traffic input sequence to an independent variable token.

[0058] Traditional Transformer-based time series modeling methods typically organize inputs by time step, treating traffic observations from all regions at the same time step as a single time token for modeling. However, for cellular network traffic data, different spatial regions often exhibit significant differences in functional attributes and statistical distributions. Organizing inputs solely by time step may fail to adequately preserve the independent semantic representation of traffic patterns across different spatial regions.

[0059] In this embodiment, the input historical traffic input sequence is assumed to be: Where T represents the length of the historical time step, Indicates the number of grid cells.

[0060] First, the input historical traffic sequence is transposed to obtain the transposed traffic sequence matrix:

[0061]

[0062] in, Represents the transposed historical flow sequence matrix, its i.e. The line represents the first Grid cells in continuous Flow sequence within a historical time step; superscript This indicates the transpose operation.

[0063] Then, a linear mapping is used to transform each length of... Historical traffic sequence mapping to In the dimensional feature space, we obtain the variable token representation matrix:

[0064]

[0065] in, This represents the variable token representation matrix obtained after inverted embedding, and its first... The line represents the first The variable token corresponding to each grid cell; Represents a learnable linear projection matrix; It represents a learnable bias vector and is broadcast along the grid cell dimension during computation; This represents the feature dimension of the variable token.

[0066] In this embodiment, after the above mapping, the complete historical traffic sequence of each grid cell is represented as an independent high-dimensional variable token, thereby providing input feature representation for subsequent spatial dependency modeling while maintaining the semantic independence of traffic in different regions.

[0067] S2-2: Introduce spatial channel embedding for each grid cell to provide a fixed spatial identity, enabling the model to learn the relatively stable spatial structural features of different grid cells and enhance the discriminative power of region representation.

[0068] In this embodiment, relying solely on variable tokens makes it difficult to explicitly distinguish the identity attributes of different spatial regions. Therefore, a learnable spatial channel embedding matrix is ​​introduced. :

[0069]

[0070] in, Indicates the number of grid cells. Indicates the feature dimension.

[0071] For the Each grid cell, and its corresponding spatial channel embedding vector, is represented as follows:

[0072]

[0073] in, This indicates the first value retrieved from the spatial channel embedding matrix. Spatial channel embedding vector corresponding to each grid cell.

[0074] Arrange the spatial channel embedding vectors of all grid cells in grid index order to obtain the spatial channel embedding representation matrix:

[0075]

[0076] in, The spatial channel embedding representation matrix represents all grid cells; Indicates the first Spatial channel embedding vector of each grid cell; Indicates the number of grid cells; Indicates the feature dimension.

[0077] In this embodiment, by introducing spatial channel embedding, different grid cells can obtain independent spatial identity representations, thereby enhancing the model's ability to express the long-term spatial attribute differences in different regions.

[0078] S2-3: Embed the time phase corresponding to the historical flow input sequence to explicitly characterize the periodic stage of the current flow sequence, thereby enhancing the model's ability to perceive long-term periodic evolution patterns.

[0079] Cellular network traffic typically exhibits significant periodic patterns, such as day-night cycles and differences between weekdays and non-weekdays. To enhance the model's ability to represent periodic patterns, this invention introduces a temporal phase embedding mechanism. Since all grid cells in the same input sample share the same temporal phase, this invention determines the temporal phase based on the last historical time step in the historical traffic input sequence and extends this temporal phase embedding to all grid cells.

[0080] In this embodiment, the period length of cellular network traffic is set to be... Construct a learnable temporal phase embedding matrix :

[0081]

[0082] in, Indicates the period length. Indicates the feature dimension.

[0083] For the last historical time step t of the input window, its corresponding time phase index is defined as:

[0084]

[0085] in, This represents the modulo operation. Indicates time step The corresponding phase index within the said period, and .

[0086] Based on time phase index The corresponding time phase embedding vector is obtained by looking up the table:

[0087]

[0088] in, This indicates the last historical time step in the input window. The corresponding time phase embedding vector; Indicates based on time phase index Embedded matrix in time phase The time phase embedding vector obtained from the query.

[0089] Since all grid cells in the same input sample are in the same time phase, the time phase embedding vector is copied along the grid cell dimension to obtain the time phase embedding representation matrix:

[0090]

[0091] in, The time-phase embedding representation matrix contains a total of They are the same time phase embedding vectors.

[0092] In this embodiment, by introducing time phase embedding, the model can explicitly perceive the periodic stage of the current input flow sequence, thereby enhancing the ability to model the periodic flow change pattern.

[0093] S2-4: Construct a spatial channel-temporal phase joint embedding for each grid cell to characterize the differentiated flow patterns of a specific spatial region under different periodic stages, thereby enhancing the model's ability to express the coupling relationship between regional flow patterns and periodic states.

[0094] In this embodiment, spatial channel embedding is used to describe the spatial properties of different grid cells, and temporal phase embedding is used to describe the periodic stage of the current input flow sequence. However, when modeled independently, these two methods are insufficient to fully express the coupling relationship between spatial regions and temporal phases. Therefore, this invention further constructs a learnable spatial channel-temporal phase joint embedding matrix. :

[0095]

[0096] For the Each grid cell and the last historical time step of the input window Corresponding time phase index The corresponding spatial channel-temporal phase joint embedding vector is expressed as:

[0097]

[0098] in, This represents the first value retrieved from the spatial channel-temporal phase joint embedding matrix. Each grid cell in time phase index The spatial channel-temporal phase joint embedding vector below; Indicates the first Each grid cell in time phase index The spatial channel-temporal phase joint embedding vector.

[0099] Arrange the spatial channel-time phase joint embedding vectors of all grid cells in the given time phase according to their grid indices to obtain the spatial channel-time phase joint embedding representation matrix:

[0100]

[0101] in, The representation matrix represents the joint embedding of spatial channels and time phases.

[0102] In this embodiment, by using the spatial channel-temporal phase joint embedding representation matrix, different grid cells can obtain differentiated representations under different time phases, thereby enhancing the model's ability to model the coupling relationship between regional flow patterns and periodic states.

[0103] S2-5: The variable token representation matrix, spatial channel embedding representation matrix, temporal phase embedding representation matrix, and spatial channel-temporal phase joint embedding representation matrix are fused to obtain the initial spatial hidden feature representation of the input global spatial Transformer:

[0104]

[0105] in, This represents the initial spatial hidden feature representation of the input global space Transformer; This represents the variable token representation matrix obtained by inverted embedding; The spatial channel embedding representation matrix; The time phase embedding representation matrix; The representation matrix represents the joint embedding of spatial channels and time phases.

[0106] In this step, through the above fusion method, the model can simultaneously obtain the historical flow change characteristics, spatial identity characteristics, temporal phase characteristics, and coupling characteristics between spatial regions and temporal phases of each grid cell, thereby forming an input representation with rich spatiotemporal semantics, providing a feature foundation for subsequent global spatial dependency modeling.

[0107] In this embodiment, the initial spatial hiding feature representation obtained in steps S2-5 is used. Input a global spatial Transformer. The global spatial Transformer models the relationships between different spatial grids through a multi-head self-attention mechanism, and can adaptively learn complex spatial dependency information within the urban area without the need for a preset fixed spatial topological adjacency matrix.

[0108] S2-6: For the first Layer Transformer encoding block, whose input is First, the query matrix, key matrix, and value matrix are obtained through linear projection.

[0109] For the One point of focus is:

[0110]

[0111] in, Indicates the first A query matrix with attention heads; Indicates the first The key matrix of each attention head; Indicates the first A matrix of values ​​for each attention head; For the first Input space hidden feature representation of a Transformer coded block; and The learnable parameter matrix; ,in This indicates the number of heads of attention.

[0112] S2-7: Calculate the spatial attention matrix using a scaled dot product attention mechanism:

[0113]

[0114] in, Indicates the first The spatial attention matrix corresponding to each attention head has elements used to characterize the spatial correlation strength between different grid cells; The feature dimension of a single attention head is represented by the following condition: ; This is a scaling factor used to prevent the dot product result from becoming too large and affecting the stability of the attention weight calculation; This represents the normalized exponential function.

[0115] S2-8: Weighted aggregation of the value matrix based on the spatial attention matrix to obtain the output of each attention head, and concatenation and linear mapping of the outputs of multiple attention heads to obtain multi-head self-attention output.

[0116] In this embodiment, the first The output of each attention head is represented as:

[0117]

[0118] in, Indicates the first The first layer of Transformer coding block Spatial information representation obtained by aggregating attention heads.

[0119] Furthermore, the outputs of all attention heads are concatenated along the feature dimension, and the multi-head self-attention output is obtained through a linear mapping:

[0120]

[0121] in, Indicates the first Multi-head self-attention output in a layered Transformer coded block; This represents a concatenation operation along the feature dimension; This represents the learnable output projection matrix.

[0122] S2-9: Combine the multi-head self-attention output with the first... The input spatial hidden feature representation of the Transformer coding block is subjected to residual connections and layer normalization to obtain an intermediate spatial hidden feature representation. This intermediate spatial hidden feature representation is then input into a feedforward network and subjected to residual connections and layer normalization again to obtain the... Output spatial feature representation of the Transformer coded block:

[0123]

[0124]

[0125] in, This represents the hidden feature representation of the intermediate space obtained after residual connection and layer normalization. Presentation layer normalization operation, Indicates a feedforward network; Indicates the first The output spatial hidden feature representation of the Transformer coded block.

[0126] S2-10: Using the output of the previous Transformer encoding block as the input of the next Transformer encoding block, repeat the above multi-head self-attention, residual connection, layer normalization, and feedforward network operations. After L layers of Transformer encoding blocks, the final spatial hidden feature representation containing global spatial dependency information is obtained:

[0127]

[0128] go through The final spatial hidden feature representation is obtained after encoding blocks of Transformer layers:

[0129]

[0130] in, This represents a Transformer coding block consisting of multi-head self-attention, residual connections, layer normalization, and a feedforward network in the l-th layer. This represents the input space hidden feature representation of the l-th layer Transformer coded block; This represents the output spatial hidden feature representation of the l-th layer Transformer coded block; Indicates the process The final spatial hidden feature representation obtained after the Transformer encoding blocks; Indicates the number of layers in the Transformer encoding block.

[0131] S2-11: Map the final spatial hidden feature representation output by the global spatial Transformer to the prediction time step dimension to obtain the spatial prediction representation:

[0132]

[0133] in, The representation space prediction is used to characterize the future continuous representation obtained from the global spatial Transformer. Spatial prediction results for each time step; This represents a learnable spatial prediction mapping matrix. It represents a learnable bias vector and is broadcast along the grid cell dimension during computation; This indicates the number of prediction steps.

[0134] S3: Divide the historical traffic input sequence into multiple time segments and input them into a segmented cyclic encoding and decoding structure to obtain a time prediction representation.

[0135] In this embodiment, the global spatial Transformer is mainly used to model the global spatial dependencies between different grid cells, while the temporal evolution of cellular network traffic is further characterized by a segmented cyclic encoding and decoding structure. To improve the prediction capability of long sequences and alleviate the error accumulation problem in recursive prediction, this invention performs segmented modeling of historical traffic input sequences and adopts a parallel multi-step prediction mechanism in the decoding stage to generate prediction results for multiple future time segments.

[0136] S3-1: Normalize the historical traffic input sequence to obtain a normalized traffic sequence.

[0137] In this embodiment, for the first Each grid cell, from the historical flow input sequence Extract the corresponding historical traffic sequence. To facilitate subsequent segmented modeling, the time steps within this historical window are renumbered. ,get:

[0138]

[0139] in,

[0140]

[0141] Indicates the first The length of each grid cell within the input window is... Historical flow sequences; Indicates the first [number]th [item] in the input window Flow observations at each historical time step; This represents the flow observation value at the last historical time step within the input window, corresponding to the global time step. Observations .

[0142] To reduce the impact of differences in flow levels between different samples on model training, this invention uses a normalization process by subtracting the observation value from the last historical time step:

[0143]

[0144] in, Indicates the first The historical flow sequence is normalized to individual grid cells. This operation means subtracting the last historical observation from the flow value at each historical time step within the input window. .

[0145] S3-2: Divide the normalized historical flow sequence into several consecutive time segments and map them to a high-dimensional feature space to obtain the segment embedding sequence.

[0146] In this embodiment, the length is Normalized historical flow sequences according to fragment length Divide into, and obtain A continuous time segment, For the first The sequence of segments of a grid cell is represented as follows:

[0147]

[0148] in, Indicates the first Each grid cell is a set of historical input segments obtained by dividing a normalized historical flow sequence; Indicates the first The first grid cell of the first grid cell A historical input segment, , where n represents the total number of historical input segments; w represents the segment length.

[0149] Subsequently, each historical input segment is mapped to a high-dimensional feature space to obtain the segment embedding sequence:

[0150]

[0151] in, Indicates the first The segment embedding sequence corresponding to each grid cell; Represents the learnable segment mapping matrix; It represents a learnable bias vector, which is broadcast along the time slice dimension during computation; Indicates the dimension of fragment embedding; This represents a non-linear activation function.

[0152] S3-3: Input the fragment embedding sequence into the GRU encoder for recursive modeling to obtain the temporal hidden representation of the historical traffic sequence.

[0153] In this embodiment, for the first The GRU encoding process for each historical input segment is as follows:

[0154]

[0155]

[0156]

[0157]

[0158] in, Indicates the first The first grid cell The embedding vector of each historical input segment, i.e. The OK; Indicates an update to the door; Indicates that the door is being reset; Indicates the first Candidate hidden states corresponding to each historical input segment; Indicates that the first step has been completed. The hidden state obtained after a historical input segment; as well as All are learnable parameters; This represents the Sigmoid activation function; ⨀ represents the hyperbolic tangent activation function; ⨀ represents element-wise multiplication.

[0159] Through the The GRU encoder recursively models the historical input segments to obtain the final hidden state:

[0160]

[0161] in, Indicates the first A temporal hidden representation of historical flow sequences for each grid cell. By employing a segmented modeling approach, long-term temporal dependencies can be captured while reducing recursion depth.

[0162] S3-4: Construct the relative temporal position embedding matrix and spatial position embedding matrix for parallel decoding, and concatenate them to obtain the final position embedding matrix.

[0163] Traditional recursive multi-step prediction typically uses the prediction result of the previous step as the input for the next step, which can easily lead to the gradual propagation and accumulation of errors as the prediction step size increases. To alleviate this problem, this invention adopts a parallel multi-step prediction mechanism. In the decoding stage, the prediction value of the previous step is no longer used as the subsequent input. Instead, prediction results for multiple future time segments are generated based on the encoder output and position embedding.

[0164] Let the prediction step size be... The predicted segment length is Then the number of future predicted segments for:

[0165]

[0166] To distinguish the temporal order of different future prediction segments, a learnable relative temporal position embedding matrix is ​​introduced:

[0167]

[0168] in, Represents the embedding matrix of relative time position; Indicates the first The relative temporal position embedding vector of each future predicted segment, where rp represents the relative position. .

[0169] Simultaneously, to describe the positional structure of mesh elements in two-dimensional space, row embedding and column embedding are introduced. For the first... Each grid cell, its two-dimensional spatial location code is represented as follows:

[0170]

[0171] in, This represents the two-dimensional spatial location code of the j-th grid cell; and They represent the first Row embedding and column embedding corresponding to each grid cell This represents a multilayer perceptron.

[0172] Will Replicating along the dimensions of future prediction segments Next, we obtain the spatial location embedding matrix:

[0173]

[0174] in, Indicates the first The spatial location embedding matrix of each grid cell; This indicates the copy function.

[0175] In this embodiment, the relative time position embedding matrix and the spatial position embedding matrix are concatenated along the feature dimension to obtain the first... The final position embedding matrix of each grid cell:

[0176]

[0177] in, Indicates the first The final position embedding matrix corresponding to each grid cell, Concat This represents a concatenation operation along the feature dimension; Indicates the first The final position embedding vector corresponding to each future predicted segment.

[0178] S3-5: Combine the temporal hidden representation output by the encoder with the final position embedding matrix and input it into the GRU decoder to obtain the set of decoded hidden states for the future predicted segment.

[0179] In this embodiment, for the first The first grid cell and the first Given a future prediction segment, construct the decoder input:

[0180]

[0181] in, Indicates the first The decoding input vector for each future predicted segment; Indicates the first Temporal hidden representation of historical traffic sequences for each grid cell; Indicates the first The final position embedding vector corresponding to each future predicted segment .

[0182] Then, the decoding inputs of all future prediction segments are arranged in chronological order to obtain:

[0183]

[0184] in, Indicates the first The set of decoding inputs for each grid cell;

[0185] Finally, The GRU decoder, with shared parameters as input, obtains the decoded hidden state of the future predicted segment:

[0186]

[0187] in, Indicates the first The first grid cell The decoded hidden state corresponding to each future predicted segment This represents the initial hidden state of the GRU decoder; Indicates the first The first grid cell -1 Decoding hidden state corresponding to a future predicted segment.

[0188] The decoded hidden states of all future predicted segments are arranged in order to form a set of decoded hidden states:

[0189]

[0190] in, Indicates the first The set of decoded hidden states for each grid cell.

[0191] S3-6: Map the features of the decoded hidden state set to the prediction value space to obtain the set of future prediction segments, and then perform scale recovery after splicing the set of future prediction segments in chronological order to obtain the time prediction representation.

[0192] In this embodiment, the set of decoded hidden states is processed... The layer undergoes regularization and uses a linear mapping to map the decoded hidden state of each future prediction segment to a linear sequence of length 1. Predicted fragment:

[0193]

[0194] in, Indicates the first Piecewise prediction results for each grid cell at the normalized scale; This represents the learnable output mapping matrix; This represents a learnable bias vector.

[0195] Will Represented as multiple predicted segments by row:

[0196]

[0197] in, Indicates the first Normalized prediction results corresponding to each future prediction segment .

[0198] Concatenate all predicted segments in chronological order to obtain the [predicted segment]. Normalized multi-step prediction results for individual grid cells:

[0199]

[0200] in, Indicates the first Normalized multistep prediction results for each grid cell over the next k time steps; Indicates a splicing operation; , where m represents the number of future predicted segments.

[0201] After obtaining the normalized prediction results, the original flow scale is restored by adding back the last historical observation:

[0202]

[0203] in, Indicates the first Future at a normalized scale for individual grid cells The prediction results at each time step This indicates the prediction result after restoring the original scale. This represents the flow observation value at the last historical time step within the input window.

[0204] Finally, the prediction results of all grid cells are arranged in grid index order to obtain the time prediction representation:

[0205]

[0206] in, This represents the time prediction representation obtained from the segmented cyclic encoding and decoding structure; This indicates the number of grid cells; k indicates the prediction step size.

[0207] Compared with traditional recursive prediction methods, the parallel decoding mechanism described in this step avoids repeatedly inputting the prediction results of the previous step into subsequent prediction steps, thereby reducing the gradual propagation and accumulation of errors within the prediction window and improving the stability and accuracy of multi-step cellular network traffic prediction.

[0208] S4: Fusion Spatial Prediction Representation and time prediction The final traffic prediction result is obtained.

[0209] In this embodiment, the spatial modeling module and the temporal modeling module characterize the spatiotemporal dependencies of cellular network traffic from different perspectives. Spatial prediction representation It mainly reflects the global spatial correlation information between different grid cells, and the time prediction representation It mainly reflects the temporal evolution of each grid cell. To fully integrate the two types of prediction information, this invention introduces a gating fusion mechanism to perform weighted fusion of spatial and temporal prediction representations.

[0210] Specifically, define learnable gating parameters. And constrain it to the (0,1) interval using the Sigmoid function:

[0211]

[0212] in, Indicates the gating fusion weights, Indicates learnable parameters, This represents the Sigmoid activation function.

[0213] The final traffic prediction result is expressed as follows:

[0214]

[0215] in, This represents the final multi-step flow prediction result; The spatial prediction representation output from step S2; The time prediction representation is the output of step S3; This indicates the number of grid cells; k indicates the prediction step size.

[0216] In this embodiment, through the gating fusion mechanism, the present invention can adjust the contribution ratio of spatial prediction information and temporal prediction information in the final prediction result according to the fusion weights learned during the training process, thereby improving the model's ability to represent complex spatiotemporal traffic patterns and improving the accuracy and stability of multi-step cellular network traffic prediction.

[0217] In this embodiment, to verify the predictive effectiveness of the present invention, experimental verification was conducted using three open-source real-world cellular network traffic datasets from both domestic and international sources: the Chinese urban cellular network traffic datasets MobileBJ and MobileNJ, and the Italian Milan urban traffic dataset Milan. These datasets exhibit significant differences in spatial scale, temporal coverage, and traffic dynamics, enabling them to be used to evaluate the predictive performance and robustness of the method under different scenarios.

[0218] (1) China City Traffic Dataset (including Beijing Dataset MobileBJ and Nanjing Dataset MobileNJ). Spatially, both datasets use a regular grid to divide the urban areas. The Beijing area (MobileBJ) is divided into 28×24 grid units, and the Nanjing area (MobileNJ) is divided into 20×28 grid units. Temporally, both datasets use a 1-hour time granularity for traffic statistics. The time coverage of MobileBJ is from October 25, 2021 to November 21, 2021, and the time coverage of MobileNJ is from February 2, 2021 to February 22, 2021.

[0219] (2) Milan Dataset: This dataset comes from the Telecom Italia Big Data Challenge and records the intensity of multi-service communication activities in Milan over several consecutive days at the grid level. The dataset divides Milan into 100×100 regular grid cells. The original time sampling interval is 10 minutes. The dataset includes anonymized service metrics such as SMS in, SMS out, call in, call out, and Internet services. In the experiment, the Milan dataset was aggregated from the original 10-minute time granularity to a 1-hour time granularity.

[0220] In this embodiment, the experiment uses historical cellular network traffic from the past 24 time steps as input to predict traffic changes over the next 12 time steps. During model training, the batch size is set to 32, the number of training epochs is set to 300, the optimizer is Adam, and the loss function is mean squared error. In the spatial modeling module, the number of stacked layers of the Transformer encoder is set to 4, the number of attention heads in the multi-head attention mechanism is set to 4, and the hidden dimension is set to 128; in the temporal modeling module, the hidden dimension is also set to 128, and the time segment length is set to 6. Specific experimental parameter configurations are shown in Table 1.

[0221] Table 1. Experimental parameter settings

[0222]

[0223] In this embodiment, to verify the technical effect of the present invention, Tables 2 and 3 respectively show the multi-step prediction performance of the GTSRNet model proposed in this invention and six comparative models on multiple real cellular network traffic datasets.

[0224] Table 2 shows the experimental results of multi-step traffic prediction on the Milan cellular network dataset, including three service scenarios: SMS, Call, and Internet. Table 3 shows the experimental results of multi-step traffic prediction on the MobileBJ and MobileNJ cellular network datasets.

[0225] Table 2. Experimental results of multi-step traffic prediction on the Milan Cellular Network dataset.

[0226]

[0227] Table 3. Experimental Results of Multistep Traffic Prediction in Cellular Networks in Beijing and Nanjing

[0228]

[0229] Experimental results show that on the Milan, MobileBJ, and MobileNJ cellular network traffic datasets, GTSRNet achieved the minimum prediction error in both RMSE and MAE evaluation metrics, verifying that the method of this invention has high prediction accuracy and good stability.

[0230] In terms of overall prediction accuracy, GTSRNet demonstrates superior prediction performance. Taking the SMS service scenario of the Milan dataset as an example, GTSRNet's Overall RMSE is 99.552, which is about 7.9% lower than the second-best TimeMixer model. In the Internet service scenario with more complex traffic fluctuations, GTSRNet's Overall RMSE is about 3.7% lower than TimeMixer. These results indicate that the method of this invention has good adaptability in handling highly volatile and nonlinear traffic sequences, and can more accurately capture complex spatiotemporal evolution patterns.

[0231] Experimental results for different prediction step lengths show that the prediction errors of all models generally increase as the prediction time increases from Step 4 to Step 12. However, GTSRNet exhibits stronger stability in multi-step prediction tasks. Taking the MobileNJ dataset as an example, when the prediction step length increases, the error growth of the traditional statistical model ARIMA and the deep learning model STID is more significant, while GTSRNet maintains a lower error level, with an RMSE of only 0.176 in Step 12. This result indicates that the method of this invention helps to alleviate the error accumulation problem in the multi-step prediction process, thereby maintaining better prediction stability.

[0232] Figure 3The results show the MAE (Modulation Error) of different models as a function of prediction step size in the Internet service scenario of the Milan cellular network dataset. With increasing prediction step size, the prediction error of all models shows an upward trend, indicating that long prediction step size tasks face greater uncertainty and modeling difficulty. Compared with other comparative models, the GTSRNet proposed in this invention achieves the minimum MAE at all prediction step sizes, and the error growth is relatively small. This demonstrates that the proposed spatiotemporal modeling mechanism and parallel multi-step prediction mechanism can effectively alleviate the error accumulation problem in multi-step prediction, thereby maintaining good stability over long prediction step sizes.

[0233] Table 4 shows the comparison results of each model in terms of training time, inference time, and parameter size. It can be seen that there is a certain trade-off between prediction performance and computational cost among the different models. Due to its more complex spatiotemporal attention mechanism, ST-Tran has higher training time, inference time, and parameter size. In contrast, the GTSRNet proposed in this invention has lower training and inference times than ST-Tran and AHSTGNN, while achieving better prediction performance. Although GTSRNet has a larger parameter size than lightweight models such as STID and TimeMixer, combined with the aforementioned prediction results, it can be seen that GTSRNet achieves higher prediction accuracy while maintaining acceptable computational cost, indicating that the designed global spatial Transformer and piecewise cyclic time modeling structure can effectively balance model accuracy and computational efficiency.

[0234] Table 4. Comparison of computational efficiency for multi-step traffic prediction in cellular networks using different models

[0235]

[0236] In this embodiment, a series of ablation experiments were designed to evaluate the effectiveness of each component module. Figure 4 , Figure 5 and Figure 6 Ablation experiments on different cellular network traffic datasets are presented. Specifically, Figure 4 The results of RMSE ablation comparison for the Milan cellular network traffic dataset. Figure 5 The results of MAE ablation comparison for the Milan cellular network traffic dataset. Figure 6This section presents RMSE ablation comparison results for cellular network traffic datasets from Beijing and Nanjing. Here, w / o S indicates removal of the global spatial Transformer, w / o STHE indicates removal of the hybrid spatiotemporal embedding layer, w / o T indicates removal of the piecewise cyclic time modeling structure, w / o G indicates removal of the gated fusion mechanism, w / o Seg indicates removal of the time series segmentation mechanism, and GTSRNet represents the complete model. By comparing the performance changes of different variants, the contribution of each component module to the model's predictive ability can be analyzed.

[0237] Experimental results show that the full model GTSRNet outperforms its corresponding ablation variants across all evaluation metrics, validating the effectiveness of the proposed framework. Removing the global spatial Transformer (w / o S) or the piecewise cyclic temporal modeling structure (w / o T) leads to significant performance degradation, indicating that both spatial dependency modeling and temporal evolution modeling play crucial roles in multi-step traffic prediction in cellular networks. The results with / o STHE further demonstrate that the hybrid spatiotemporal embedding mechanism enhances the model's ability to represent spatial heterogeneity and temporal periodicity. Compared to the full model, both w / o G and w / o Seg exhibit varying degrees of performance degradation, indicating that the gated fusion mechanism can more effectively integrate spatial and temporal prediction information, while the piecewise mechanism helps improve the model's ability to model long-sequence dependencies.

[0238] Figure 7 The results are visualized for 12 consecutive prediction steps on a single day from the MobileNJ cellular network traffic dataset. Overall, GTSRNet effectively tracks the actual traffic trend, exhibiting high fitting accuracy near peaks, troughs, and trend inflection points. As the prediction step size increases, some comparative models gradually deviate from the true values, while GTSRNet maintains stable prediction performance. Especially in the later prediction stages (t+7 to t+12), GTSRNet can still accurately depict the traffic decline trend, indicating that the proposed model possesses strong long-term dependency modeling capabilities and multi-step prediction stability, effectively mitigating performance degradation caused by error accumulation.

[0239] To further evaluate the predictive stability of the model in complex spatiotemporal scenarios, this invention selects the 12th prediction step, where the error accumulation phenomenon is more significant, for spatial visualization analysis. Figure 8 The results comparing GTSRNet and TimeMixer on real traffic distribution and absolute error spatial projection are presented.

[0240] From the spatial distribution of the actual and predicted values, GTSRNet maintains good spatial consistency with the actual flow field, demonstrating good characterization accuracy in both high and low flow regions. In contrast, TimeMixer exhibits significant deviations in some high-load regions and generates large prediction errors in several local hotspots. Further observation of the spatial projection of the absolute error reveals that GTSRNet's error distribution is more uniform, with a lower overall error level, indicating its ability to more effectively model complex spatiotemporal dependencies.

[0241] The above results demonstrate that the model proposed in this invention can not only accurately capture the global spatial distribution characteristics of cellular network traffic, but also maintain good spatial prediction stability in scenarios with long prediction step lengths, thereby effectively alleviating the performance degradation problem caused by error propagation during multi-step prediction.

[0242] Those skilled in the art will understand that the above embodiments are specific examples of implementing the present invention, and in practical applications, various changes in form and detail may be made without departing from the spirit and scope of the present invention.

Claims

1. A multi-step traffic prediction method for cellular networks, characterized in that, Specifically, the following steps are included: S1: Obtain historical cellular network traffic sequences from the server and preprocess them to obtain historical traffic input sequences; S2: After the historical traffic input sequence is mixed with spatiotemporal embedding, it is input into the global spatial Transformer, and the spatial prediction representation is obtained by using the multi-head self-attention mechanism; S3: Divide the historical traffic input sequence into multiple time segments and input them into a segmented cyclic encoding and decoding structure to obtain a time prediction representation; S4: The final traffic prediction result is obtained by fusing the spatial prediction representation and the temporal prediction representation.

2. The multi-step traffic prediction method for cellular networks as described in claim 1, characterized in that, S1 specifically includes the following steps: S1-1: Obtain cellular network traffic data for the target area within a continuous historical time period from the server, and divide the target area into... The system consists of a regular grid cell, and the two-dimensional flow graph of each time step is unfolded into a one-dimensional flow vector according to the grid order. S1-2: Select continuous One-dimensional flow vectors at each historical time step are stacked in chronological order to obtain the historical flow input sequence.

3. The multi-step traffic prediction method for cellular networks as described in claim 1, characterized in that, S2 specifically includes the following steps: S2-1: The inverted embedding method is used to map the historical flow sequence of each grid cell in the historical flow input sequence to the corresponding variable token representation matrix; S2-2: Introduce spatial channel embedding for each grid cell to obtain the spatial channel embedding representation matrix; S2-3: Embed the time phase corresponding to the historical flow input sequence to obtain the time phase embedding representation matrix; S2-4: Construct a spatial channel-temporal phase joint embedding for each grid cell to obtain the spatial channel-temporal phase joint embedding representation matrix; S2-5: The variable token representation matrix, spatial channel embedding representation matrix, temporal phase embedding representation matrix, and spatial channel-temporal phase joint embedding representation matrix are fused to obtain the initial spatial hidden feature representation of the input global space Transformer; S2-6: For the first The input to the Transformer encoding block is first obtained by linear projection to obtain the query matrix, key matrix, and value matrix respectively; S2-7: Calculate the spatial attention matrix using a scaled dot product attention mechanism; S2-8: Weighted aggregation of the value matrix is ​​performed based on the spatial attention matrix to obtain the output of each attention head, and the outputs of multiple attention heads are concatenated and linearly mapped to obtain the multi-head self-attention output; S2-9: Combine the multi-head self-attention output with the first... The input spatial hidden feature representation of the Transformer coding block is subjected to residual connections and layer normalization to obtain an intermediate spatial hidden feature representation. This intermediate spatial hidden feature representation is then input into a feedforward network and subjected to residual connections and layer normalization again to obtain the... Output spatial feature representation of the layer Transformer coded block; S2-10: Use the output spatial hidden feature representation of the previous Transformer coding block as the input of the next Transformer coding block, and repeat the above operation. After passing through multiple Transformer coding blocks, the final spatial hidden feature representation containing global spatial dependency information is obtained. S2-11: Map the final spatial hidden feature representation to the prediction time step dimension to obtain the spatial prediction representation.

4. The multi-step traffic prediction method for cellular networks as described in claim 3, characterized in that, In S2-6, for the first Input of the layer Transformer coding block The query matrix, key matrix, and value matrix corresponding to the r-th attention head are as follows: in, express The first layer of Transformer coding block A query matrix with attention heads; Indicates the first The key matrix of each attention head; Indicates the first A matrix of values ​​for each attention head; For the first Spatial hidden feature representation of the input to the layer Transformer coded block; and The learnable parameter matrix; , This indicates the number of heads of attention.

5. The multi-step traffic prediction method for cellular networks as described in claim 3, characterized in that, In S2-7, the spatial attention matrix is: in, Indicates the first The spatial attention matrix corresponding to each attention head; Indicates the first A query matrix with attention heads; Indicates the first The key matrix of each attention head; The feature dimension of a single attention head; Represents the normalized exponential function; superscript This indicates the transpose operation.

6. The multi-step traffic prediction method for cellular networks as described in claim 3, characterized in that, In S2-8, the first The outputs of individual attention heads and multi-head self-attention heads are as follows: in, Indicates the first The first in the layer The output of each attention head; Indicates the first The spatial attention matrix corresponding to each attention head; Indicates the first A matrix of values ​​for each attention head; Indicates the first Multi-head self-attention output of the layer; This represents a concatenation operation along the feature dimension; This represents the learnable output projection matrix; Indicates the number of heads of attention. .

7. The multi-step traffic prediction method for cellular networks as described in claim 3, characterized in that, In S2-11, spatial prediction is expressed as: in, Spatial prediction representation; This represents a learnable spatial prediction mapping matrix. Represents a learnable bias vector; Indicates the process The final spatial hidden feature representation obtained after the Transformer encoding block; This indicates the transpose operation.

8. The multi-step traffic prediction method for cellular networks as described in claim 1, characterized in that, S3 specifically includes the following steps: S3-1: Normalize the historical flow input sequence to obtain a normalized flow sequence; S3-2: Divide the normalized flow sequence into several consecutive time segments and map each time segment to a high-dimensional feature space to obtain the segment embedding sequence; S3-3: Input the fragment embedding sequence into the GRU encoder for recursive modeling to obtain the temporal hidden representation of the historical traffic sequence; S3-4: Construct the relative temporal position embedding matrix and spatial position embedding matrix for parallel decoding, and concatenate them to obtain the final position embedding matrix; S3-5: Combine the temporal hidden representation output by the encoder with the final position embedding matrix and input it into the GRU decoder to obtain the set of decoded hidden states for the future predicted segment; S3-6: Map the features of the decoded hidden state set to the prediction value space to obtain the set of future prediction segments, and then perform scale recovery after splicing the set of future prediction segments in chronological order to obtain the time prediction representation.

9. A multi-step traffic prediction method for cellular networks as described in claim 8, characterized in that, In S3-6, the time prediction is expressed as: in, Indicates time prediction; Indicates the number of grid cells; k represents the prediction step size. Indicates the number of future predicted segments; Indicates the first Time prediction results for each grid cell over the next k time steps; Indicates the first Normalized multistep prediction results for each grid cell over the next k time steps; Indicates the first Each grid cell is in the last historical time step of the input window. Flow observations; Indicates the first The grid cell in the first... Normalized prediction results in a future prediction segment .

10. A multi-step traffic prediction method for cellular networks as described in claim 1, characterized in that, In step S4, the final traffic prediction result is: in, This represents the final multi-step flow prediction result; Spatial prediction representation; Indicates time prediction; Indicates learnable gating parameters; This represents the Sigmoid activation function; This indicates the gating fusion weight.

Images