A Synchronous Spatiotemporal Multi-Step Traffic Flow Prediction Method Based on Hybrid Graph Convolution

Abstract

The present invention discloses a synchronous spatio-temporal base station traffic prediction method based on hybrid graph convolution. Aiming at the challenges of the complex mobility of users and the randomness of the network scheduling and allocation mechanism, as well as the phenomenon of local over-concentration in extracting traffic spatial features. The present invention uses a gated temporal convolutional network to finely extract three types of short-term traffic features: near-time, near-day, and near-week. At the same time, it fuses a dynamic graph convolutional network based on the spatial domain and an adaptive graph convolutional network based on the spectral domain. The dual-graph structure can effectively extract the sudden random dependence relationship between users and network scheduling. Finally, to avoid the phenomenon of local feature over-concentration in the graph convolutional network, a spatio-temporal gating mechanism is innovatively introduced to dynamically fuse temporal features and spatial features. Experimental results show that this scheme has good performance in both short-term and medium- and long-term predictions, and even at the last time step, the error can be maintained within a relatively stable range.

Description

Technical Field

[0001] The present invention relates to the technical field of network traffic analysis, and specifically to a synchronous spatio-temporal multi-step traffic prediction method based on hybrid graph convolution. Background Technique

[0002] There are multiple base station nodes in a wireless network. Different cellular base stations maintain coverage of different geographical areas. Users can connect to the cellular base station corresponding to their covered area to access the mobile network and consume cellular traffic. Due to the limited coverage range of wireless signals, the movement behavior of users between regions will bring about handover allocation between multiple cellular base stations, which results in different traffic patterns at different times for cellular base stations located at different positions. Accurate cellular traffic prediction technology helps to improve communication network management, enhances the security and stability of network performance, plays an irreplaceable role in the planning and construction of public networks, and provides strong support for the development and optimization of the network.

[0003] The daily behaviors of most users are to periodically visit relatively fixed places, which means that there is a relatively stable dynamic proximity dependence relationship when users connect to cellular base stations. Due to the convenience of modern transportation, users can spontaneously and randomly visit other distant locations. At the same time, for a better user experience, the wireless network scheduling mechanism can also allocate users to the nearest cellular base station, or a base station with fewer users, or a remote base station with stronger signals, etc. The uncertainty of this user movement behavior and the randomness of the wireless network scheduling mechanism mean that there is also a sudden random dependence relationship when users connect to cellular base stations, even for cellular base stations that are far away, which increases the difficulty of extracting spatial features.

[0004] When a graph convolutional network extracts features, due to its structure and working principle, it may be more inclined to focus on local information. This is because the graph convolutional network updates the representation of nodes by gradually aggregating the information of neighbor nodes through layer-by-layer convolution. In this process, local information is often easier to capture and emphasize, which may result in an error in spatial heterogeneity, that is, in a specific time period, the central urban area should be in a dormant state at night, but the prediction result shows that cellular traffic over-aggregates in the central urban area. If network planners allocate network resources based on these prediction results, it may lead to an excess of network resources in the central urban area and insufficient resources in suburban or rural areas, which not only wastes resources but also affects user experience and service quality. Summary of the Invention

[0005] (1) Technical Problems to be Solved

[0006] Aiming at the deficiencies of the existing technology, the present invention provides a synchronous spatio-temporal multi-step traffic prediction method based on hybrid graph convolution. This method uses a fine-grained time slicing module to extract short-term time series traffic data divided into three different time intervals, and uses a gated time series convolutional network to finely extract such short-term time series features. The present invention also fuses a dynamic graph convolutional network based on the spatial domain and an adaptive graph convolutional network based on the spectral domain to capture the dynamic proximity features and sudden random distance features in the space of traffic data respectively. At the same time, to avoid the problem that the graph convolutional network over-focuses on local features, a spatio-temporal gating mechanism is innovatively introduced to dynamically balance the fusion of time series features and spatial features.

[0007] (II) Technical solution

[0008] To achieve the above technical objectives, the present invention provides the following technical solutions: including the following steps:

[0009] Step 1: According to the periodicity and seasonal characteristics of cellular traffic in time series, the spatio-temporal data set is finely sliced and divided into three historical time steps of the same length, and then a gated time series convolutional network is used to extract the short-term time series features of traffic;

[0010] Step 2: According to the randomness and suddenness of cellular traffic in space, a dynamic graph convolutional network based on the spatial domain and an adaptive graph convolutional network based on the spectral domain are fused, and a hybrid graph structure is used to extract the spatial features of traffic;

[0011] Step 3: Aiming at the problem that the graph convolutional network over-focuses on local features, the spatial features and short-term time series features of data are dynamically weighted and fused in the synchronous spatio-temporal feature fusion module. After extracting the hybrid spatial features, the time features carrying historical data are used to dynamically correct this deviation. The formula of the synchronous spatio-temporal feature fusion module is as follows:

[0012]

[0013]

[0014] where, W t , W g , b f are all learnable parameters, H l ∈R T×N×D is both the output of the STFM module and the output of the l-th layer of the SSTHL module;

[0015] Step 4. The output of the synchronous spatio-temporal feature fusion module in Step 3 is the output of the l-th layer of the SSTHL module. A parameterized skip connection is introduced to add the output of the previous layer to the input of the current layer. The number of channels of the skip connection and the number of channels at the last time are set. After the skip connection is completed, two fully connected layers are used to implement the output layer. Finally, the predicted multi-step result is M is the time step length for predicting the future.

[0016] Preferably, Step 1 includes the following steps:

[0017] Step 1.1. The Time Slice Module (TSM) slices the spatio-temporal dataset of cellular traffic into three historical time steps of the same length according to different adjacent time intervals, including extraction by adjacent time, adjacent day, and adjacent week.

[0018] Step 1.2. The Temporal Feature Extraction Module (TFM) applies the concept of gates and consists of three aggregated gated temporal convolutional networks. Each gated component consists of two dilated causal convolutional blocks to capture the temporal trends of nodes, and then a multi-layer perceptron (MLP) is used to aggregate the three temporal features.

[0019] Preferably, Step 2 includes the following steps:

[0020] Step 2.1. The Adaptive Dynamic Graph Convolutional Network based on the Spatial Domain (ADGL) is composed of a dynamic graph attention network (GAT), a multi-head attention mechanism, and a prior graph structure that fuses geographical information and semantic information, and is used to effectively extract the dynamic neighborhood of users.

[0021] Step 2.2. The Adaptive Graph Convolutional Network based on the Spectral Domain (SAGL) is composed of an adaptive adjacency matrix generated by the random node embedding method or the singular value decomposition method, and is used to effectively extract the sudden stochastic dependence relationship between users and network scheduling.

[0022] Step 2.3. The spatial correlation of cellular traffic is associated with two graph convolution calculations. The traffic has both stable dynamic spatial neighborhood characteristics and hidden random distance spatial dependence relationships. The calculation results of these two graph structures are weighted and fused.

[0023] Preferably, in Step 3, different cellular base stations are affected by different temporal and spatial features, and the nodes have different spatio-temporal trends. After extracting the mixed spatial features, the time features carrying historical data are used to dynamically correct the bias, and a gated unit is used to fuse the temporal and spatial features of the traffic.

[0024] Preferably, four SSTHL modules are stacked in Step 3. A parameterized skip connection is introduced to add the output of the previous layer to the input of the current layer. Finally, two fully connected layers are used to output the final multi-step prediction result.

[0025] Preferably, when approaching, the cellular traffic values at adjacent previous and next time steps are the same. To predict the cellular traffic data for M time steps after the t-th moment in the future, select the cellular traffic data for the adjacent hour, i.e., T time steps, before the t-th moment as the input. Its expression is as follows:

[0026] X h =(X t-T+1 , X t-T+2 ,.., X t ), t>T

[0027] Among them, X h is the data set at the t-th moment obtained by dividing the cellular traffic data when approaching.

[0028] Preferably, among the cellular traffic with a daily cycle extracted when approaching the day, the traffic value at a fixed moment on the previous day is the same as the traffic value at the same fixed moment on the next day. To predict the cellular traffic data for M time steps after the t-th moment in the future, select the cellular traffic data at the same moment on the previous day of the t-th moment as the slicing starting point, and slice backward to obtain the cellular traffic data of T time steps as the input. Its expression is as follows:

[0029] X d =(X t-24×q+1 , X t-24×q+2 ,.., X t-24×q+T ), t>24×q

[0030] Among them, X d is the data set at the t-th moment obtained by dividing the cellular traffic data when approaching the day.

[0031] Preferably, when approaching the week, the cellular traffic with a weekly cycle ignores the differences in daily patterns between weekdays and holidays. Therefore, among the cellular traffic on a weekly basis, the traffic value at a certain fixed moment in the previous week is the same as the traffic value at a certain fixed moment in the next week. To predict the cellular traffic data for M time steps after the t-th moment in the future, select the cellular traffic data from the previous week of the t-th moment as the slicing starting point, and slice to obtain the sliding input of T time steps. Its expression is as follows:

[0032] X w =(X t-24×7×q+1 , X t-24×7×q+2 ,.., X t-24×7×q+T ), t>7×24×q

[0033] Among them, X w is the data set at the t-th moment obtained by dividing the cellular traffic data when approaching the week.

[0034] Preferably, divide the cellular traffic data into the X at the t-th moment according to approaching the moment, approaching the day, and approaching the weekh , X d and X w ∈R T×N×F , where N represents the number of nodes, F represents the feature dimension on the nodes. As time t moves backward gradually, a sliding window data set X with different historical information is obtained H , X D , X W , and its expression is as follows:

[0035] L = (t - 7×24)q - M + 1

[0036]

[0037] where L is the length of the sliding window data set.

[0038] Preferably, the sets X H , X D and X W ∈R L×T×N×F are respectively input into the local temporal - feature extraction module LTFM. At this time, it is a multi - node spatio - temporal traffic prediction problem, and only the cellular traffic feature values of multiple nodes are predicted. At this time, N > 1 and F = 1.

[0039] Compared with the prior art, the present invention provides a synchronous spatio - temporal multi - step traffic prediction method based on hybrid graph convolution, having the following beneficial effects:

[0040] The advantages of the present invention over the prior art are as follows: In view of the challenges of the complex mobility of users and the randomness of the network scheduling and allocation mechanism faced by the cellular base station traffic prediction in the network, and the local over - concentration phenomenon existing in the existing models when extracting the spatial features of cellular traffic, by proposing a hybrid spatial - feature extraction module, which fuses the dynamic graph convolution network based on the spatial domain and the adaptive graph convolution network based on the spectral domain, to capture the dynamic neighborhood features and the sudden random - distance features in the space of traffic data respectively. Also, by introducing a spatio - temporal gating mechanism to dynamically balance the fusion of temporal features and spatial features, the local over - concentration phenomenon is effectively solved. The experimental results show that the proposed scheme has high feasibility, and the present invention can achieve better performance in both short - term and medium - to long - term predictions of spatio - temporal traffic. Even at the last prediction time step, the error can be maintained within a relatively stable range. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 is the overall structure diagram of the synchronous spatio - temporal multi - step traffic prediction model based on hybrid graph convolution disclosed in the embodiments of the present invention;

[0042] Figure 2 is the schematic diagram of the fine - grained slicing module disclosed in the embodiments of the present invention;

[0043] Figure 3 It is a detailed diagram of the fine-grained temporal feature extraction module disclosed in the embodiments of the present invention;

[0044] Figure 4 It is a detailed diagram of the synchronous spatio-temporal fusion module disclosed in the embodiments of the present invention;

[0045] Figure 5 It is a detailed diagram of the output layer disclosed in the embodiments of the present invention. Detailed implementation manners

[0046] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0047] The present invention is composed of a time fine-grained slicing module (Time Slice Module, TSM) and multiple stacked SSTHL blocks. Each SSTHL block consists of four parts: a fine-grained temporal feature extraction module (Temporal Feature Module, TFM), a hybrid spatial feature extraction module (Hybrid Spatial Feature Module, HSFM), a synchronous spatio-temporal fusion module (Spatio-Temporal Fusion Module, STFM), and a final output layer module (Output Layers).

[0048] The present invention will be described below with examples:

[0049] 1. Time fine-grained slicing module

[0050] According to different adjacent time intervals, the cellular traffic spatio-temporal data set is finely sliced and divided into three historical time steps of the same length according to adjacent hours, adjacent days, and adjacent weeks. Assuming the size of the historical sliding window is T steps, the sampling frequency per day is q, and the cellular traffic values for the next M time steps are predicted. For the multi-step prediction problem, M>1. The process is as follows:

[0051] (1) Extraction by adjacent hour

[0052] The cellular traffic values at adjacent front and back time steps of the cellular traffic may be the same. To predict the cellular traffic data for the next M time steps after the t-th moment, the cellular traffic data of the adjacent hours before the t-th moment, that is, T steps, is selected as the input:

[0053] X h =(X t-T+1,X t-T+2 ,..,X t ),t>T

[0054] (2) Extract by adjacent days

[0055] In the cellular traffic with a daily cycle, since human daily routines follow a regular pattern, the traffic values at a fixed time on the previous day are likely to be the same as those at the same fixed time on the next day. To predict the cellular traffic data at M time steps after time t in the future, select the cellular traffic data at the same time on the previous day of time t as the starting point of the slice, and cut backward T steps of cellular traffic data as the input:

[0056] X d =(X t-24×q+1 ,X t-24×q+2 ,..,X t-24×q+T ),t>24×q

[0057] (3) Extract by adjacent weeks

[0058] The cellular traffic with a weekly cycle ignores the differences in daily patterns between weekdays and holidays. Therefore, in the cellular traffic on a weekly basis, the traffic values at a certain fixed time in the previous week are likely to be the same as those at a certain fixed time in the next week. To predict the cellular traffic data at M time steps after time t in the future, select the cellular traffic data from the previous week of time t as the starting point of the slice, and cut T time steps of sliding input as:

[0059] X w =(X t-24×7×q+1 ,X t-24×7×q+2 ,..,X t-24×7×q+T ),t>7×24×q

[0060] Divide the cellular traffic data according to adjacent hours, adjacent days, and adjacent weeks to obtain X at time t h 、X d 、X w ∈R T ×N×F . Where N represents the number of nodes and F represents the feature dimension on the nodes. As time t gradually moves backward, a sliding window data set X with different historical information is obtained H 、X D 、X W :

[0061] L=(t - 7×24)q - M + 1

[0062]

[0063]

[0064] Finally, the set X H , X D , X W ∈R L×T×N×F are respectively input into the local temporal feature extraction module LTFM. Since this is a multi-node spatio-temporal traffic prediction problem and only the eigenvalue of the cellular traffic of multiple nodes is predicted, at this time N>1 and F = 1.

[0065] 2. Fine-grained temporal feature extraction module

[0066] The fine-grained temporal feature extraction module TFM applies the gate concept. Each TFM consists of three aggregated gated temporal convolutional networks (Gated-TCN), and the three gated components are named GTCN-H, GTCN-D, and GTCN-W, which respectively extract the temporal features of the cellular traffic near the time, near the day, and near the week. The gated temporal convolutional network is an improved network structure that introduces a gating mechanism on the basis of the temporal convolutional network (TCN). It utilizes the causality of the temporal convolutional network and the characteristics of dilated convolution, and expands the receptive field of the model by stacking multiple convolutional layers, so as to capture the long-term dependence relationship in the time series data. Each gated component consists of two parts: a filter convolution layer and a gate convolution layer.

[0067] The filter convolution layer is responsible for extracting the features in the input time series data. It is implemented through a series of two-dimensional convolutional layers (nn.Conv2d) with different dilation rates. The number of input and output channels of the first convolutional layer is 32, the kernel size (kernel_size=(1,2)), the stride (stride=(1,1)), and the dilation rate is the default value; the number of input and output channels, kernel size, and stride size of the second convolutional layer are the same as the first time, but the dilation rate (dialation=(2,2)) is modified. Each filter convolution layer outputs the same number of channels as the input (i.e., dilation_channels = 32) so as to perform element-wise multiplication with the output of the gate convolution layer.

[0068] The gate convolution layer is responsible for controlling the transmission of information. Its structure is similar to that of the filter convolution layer, and it also uses a series of two-dimensional convolutional layers with different dilation rates. However, the output of the gate convolution layer will pass through a Sigmoid activation function to compress the output value to between 0 and 1 to form a gating signal. This gating signal will perform element-wise multiplication with the output of the filter convolution layer to control the transmission of information. Specifically, when the value of the gating signal is 1, it means that the information at this position is completely retained; when the value of the gating signal is 0, it means that the information at this position is completely discarded.

[0069] After these three gating components are computed in parallel, a multi-layer perceptron (MLP) aggregates the results. The output of the fine-grained temporal feature extraction module TFM at the l-th layer is

[0070] 3. Hybrid Spatial Feature Extraction Module

[0071] (1) Adaptive Dynamic Graph Learning Module ADGL Based on Spatial Domain

[0072] Cellular traffic data is CDR data generated based on grid regions. Each cellular base station covers a certain geographical area. Due to the mobility of the population and the specific geographical location of the base stations, there is a certain correlation in the cellular traffic data between grids that are relatively close. This is a static and intuitive near-neighbor spatial correlation. Intuitively, there is a certain relationship between the degree of interaction of traffic data between base stations and the adjacent geographical space distance where they are located. Generally speaking, the closer the distance, the higher the degree of spatial correlation. Considering the center points of each grid, the Euclidean distance between the grid and the target grid is calculated using the longitude and latitude values of different grid centers, and a distance-based graph adjacency matrix is generated using the Gaussian radial kernel function, i.e., the exponential decay function. This matrix is called the distance adjacency matrix.

[0073] In addition to this intuitive spatial distance correlation of geographical grids, according to the positive and negative correlations of cellular traffic, even if grids have the same spatial distance from surrounding target grids, their spatial correlations may be different. This is a hidden near-neighbor spatial correlation. The similarity function can be used to reflect the potential correlation between different grids from a semantic perspective. Using the similarity coefficient based on time series, the Pearson's adjacency matrix can be calculated using the Pearson correlation coefficient.

[0074] To fuse this intuitive near-neighbor correlation and hidden near-neighbor correlation between grids, this paper defines a newly defined nearest-neighbor adjacency matrix, denoted as the nearest-neighbor adjacency matrix. This matrix is obtained by calculating the Hadamard product of the distance adjacency matrix and the Pearson's adjacency matrix.

[0075] The present invention believes that the dynamic influence of short-term cellular traffic between grids mainly occurs between adjacent cellular base stations. Therefore, SDGL uses the dynamic graph GAT to aggregate dynamic spatial features. This kind of near-neighbor spatial feature is related to the geographical distance with intuitive near-neighborliness and also related to the similarity distance with hidden near-neighborliness. Therefore, we use the above-mentioned nearest-neighbor adjacency matrix as the prior graph structure of GAT to accelerate the weighted calculation of the model's attention weights.

[0076] The dynamic graph GAT is used to learn the attention weight coefficients between neighboring nodes and update the hidden feature settings of the nodes. The network also uses the multi-head attention mechanism to capture the semantic relationships between nodes in different learning subspaces. Multi-head attention can enrich the training of the model and reduce the time complexity through parallel computing.

[0077] In the initial initialization stage, the loss rate of the intermediate layer is set (dropout = 0.3), only the first-order neighbor information is considered so the neighbor order is set (order = 1), and the number of multi-head attentions is set (nheads = 4). Initialize the attention layer list of the first-order neighbors (depth = list(range(blocks * layers)), which is used to calculate the attention weights of the first-order neighbors. The depth of the dynamic graph stack is related to the number of GAT modules and layers and is set to 8 (blocks = 4, layers = 2). The number of input and output channels is 32, and the negative slope parameter of the attention mechanism is (alpha = 0.2).

[0078] Apply dropout to the input temporal feature x, use the attention layer list self.attentions of the first-order neighbors to calculate the output of each attention head, and concatenate them. Apply dropout to the concatenated output again. Use the self.out_att layer to perform the final attention calculation and aggregation on the concatenated output, and apply the ReLU activation function to return the final node embedding. In the final attention layer, aggregate the outputs of multiple attention heads and neighbor orders and output the final embedding result. The output of the l-th layer of ADGL is H l DG ∈R T×N×D 。

[0079] (2) Spectral-domain based Adaptive Graph Convolutional Network SAGL

[0080] A new type of adaptive adjacency matrix is used. This adaptive adjacency matrix does not require any prior graph structure and conducts end-to-end adaptive learning through stochastic gradient descent, enabling the model to obtain hidden random distance space dependencies by itself, making up for the deficiency of extracting spatial features from fixed graph structures.

[0081]

[0082] Among them, two randomly initialized learnable parameters are initialized in the way of random node embedding C represents the node embedding dimension, which is generally set to 10. Assume E1 is the original node embedding and E2 is the target node embedding. Through The multiplication calculation can measure the similarity between two embedding vectors, that is, obtain the spatial dependence weight value between the original node and the target node. In the subsequent adaptive graph convolution operation, if the learnable parameters have been initialized, the singular value decomposition method (SVD) is used to update the node identification. The role of the ReLU function is to remove the weak connection values close to zero in the adjacency matrix, and then directly obtain Rather than calculating the symmetric normalized Laplacian matrix as in the traditional graph convolution model, the adaptive graph convolution calculation process of the l-th layer SAGL module can be derived according to the above formula:

[0083] where, I N ∈R N×N is the identity matrix, Θ ∈ R D×D is the learnable graph convolution kernel. Generally, a dilated convolution layer is set (dilation_channels = 32), and the output of the l-th layer ADGL is

[0084] (3) Spatial feature fusion

[0085] The spatial correlation of cellular traffic data is related to both of the above two graph convolution calculations. Cellular traffic has both relatively stable dynamic spatial proximity characteristics and hidden random distance spatial dependence relationships. By fusing the calculation results of these two graph structures, the output of the l-th layer spatial feature extraction module HSFM is

[0086] 4. Synchronous spatio-temporal fusion module

[0087] When the graph convolution network extracts features, due to its structure and working principle, it may be more inclined to focus on local information. This is because the graph convolution network updates the node representation by gradually aggregating the information of neighbor nodes layer by layer. In this process, local information is often easier to capture and emphasized. That is, during a specific time period, the central urban area should be in a dormant state at night, but the prediction results show that cellular traffic is overly aggregated in the central urban area. To solve this problem of over-focusing on local features, the present invention proposes a gated synchronous spatio-temporal fusion module STFM to capture the spatio-temporal trends at the node level. After extracting the mixed spatial features, the time features carrying historical data are used to dynamically correct this deviation. The STFM formula is as follows:

[0088]

[0089] where, W t 、W g 、b f are all learnable parameters, H l ∈R T×N×DIt is both the output of the STFM module and the output of the l-th layer SSTHL module.

[0090] 5. Output layer module

[0091] As the SSTHL modules are stacked, the receptive field of the present invention continuously expands. In this paper, four SSTHL modules are stacked. At the bottom layer, it pays more attention to the temporal traffic characteristics of neighbors, and at the top layer, it pays more attention to long-term spatio-temporal information. The output of the above synchronous spatio-temporal fusion module STFM is the output of the l-th layer SSTHL module. To solve the problem of gradient disappearance or gradient explosion during model training, parametric skip connections are introduced, adding the output of the previous layer to the input of the current layer, and setting the number of channels of the skip connection (skip_channels = 256) and the number of channels of the last layer (end_channels = 512). After the skip connection is completed, two fully connected layers are used to implement the output layer, and the finally output predicted multi-step result is M is the time step length for predicting the future.

[0092] 6. Experiments and result analysis

[0093] 6.1 Evaluation metrics

[0094] To measure the prediction effect, the mean absolute error MAE and the root mean square error RMSE are selected as evaluation metrics. The closer the two numbers are to zero, the higher the model prediction effect.

[0095] 6.2 Experimental environment parameters

[0096] Table 1 Experimental environment parameters

[0097]

[0098]

[0099] 6.3 Hyperparameter selection and setting

[0100] This model uses pytorch to implement the prediction model. The sliding window size is default set to 6. The model is trained using the Adam optimizer, which can replace the traditional gradient descent process, adaptively adjust parameters and update the step size, accelerating the convergence process. The number of nodes is set to 900, the training learning rate is set to 0.01, the inactivation rate is set to 0.3, the number of training iteration epochs is 100, the batch size is set to 32, and the number of neurons in the hidden layer is set to 32.

[0101] Table 2 Model parameter selection

[0102]

[0103] 6.4 Experimental results and analysis

[0104] The present invention selects a public spatiotemporal cellular traffic data set, the Milan data set, for prediction. The set contains three sub-data sets, namely Short Message Service (SMS), Phone Call Service (Call), and Cellular Traffic (Internet). The Milan data set counts 100×100 cellular base station data at a time interval of 10 minutes, so the invention chooses to summarize these three traffic sub-data sets every hour. Due to limited computing power, 900 center grids with horizontal and vertical coordinates ranging from [40,69] are selected as prediction grids.

[0105] Considering multi-step prediction, the prediction results of the last six steps of the model are carefully analyzed. It can be observed from the table that the prediction effect of the present invention on the Internet data set is the worst, indicating that the model has the worst interpretability for the Internet data set. The reason may be that the self-similarity of the Internet data set is closer to 0.5, indicating that the data distribution of the data set is closer to disordered Brownian motion, with more burst traffic data, more drastic fluctuations in traffic, and greater differences in data peaks and valleys compared to SMS and CALL. Therefore, the Internet has more complex spatiotemporal characteristics. The present invention can achieve the best performance in both short-term and medium- and long-term predictions, and even the error in the last time step remains relatively stable. The performance improvement also shows that the present invention is more suitable for processing cellular traffic data with complex spatiotemporal dependencies.

[0106] Table 3 Comparison of the prediction results of the Milan spatiotemporal dataset for the next 6 steps

[0107]

[0108]

[0109] The parameter quantity of the present invention is 278070, and the training time is 3.197 (s / epoch). Overall, the present invention has significantly improved the prediction accuracy of cellular traffic, and has better performance in long-term multi-step prediction. Therefore, it is worthwhile to increase the consumption of some training time and a slightly more complex model structure to obtain better performance improvement. With the development of hardware equipment, the cost gap of complex models in training time will gradually narrow.

[0110] The present invention aims at the challenges of user mobility and complexity of network scheduling mechanism faced by cellular base station traffic prediction in wireless networks, as well as the problem of local over-concentration in the existing models when extracting spatial features of cellular traffic. In the spatiotemporal sequence of multi-node cellular traffic, a synchronous time-based base station traffic prediction method based on hybrid graph convolution is designed. The invention innovatively proposes a hybrid spatial feature extraction module to fuse dynamic graph convolutional networks and adaptive graph convolutional networks, wherein the dynamic graph convolutional network based on the spatial domain is composed of a dynamic graph GAT and a multi-head attention mechanism and a priori graph structure that integrates geographic information and semantic information, which can effectively extract the dynamic neighbor of the user, and the adaptive graph convolutional network based on the spectral domain is composed of an adaptive adjacency matrix generated based on a random node embedding method or a singular value decomposition method (SVD), which can effectively extract the sudden random dependency between users and network scheduling. Finally, in order to reduce the problem of excessive attention of the graph convolutional network to local connections, a synchronous spatiotemporal feature fusion module is proposed to dynamically weighted fuse the temporal features and spatial features of the traffic. Experimental results show that the scheme has high feasibility.

[0111] It should be understood that the specific order or hierarchy of steps in the disclosed process is an example of an exemplary method. Based on design preferences, it should be understood that the specific order or hierarchy of steps in the process can be rearranged without departing from the scope of protection of the present disclosure. The attached method claims present the elements of the various steps in an exemplary order and are not intended to be limited to the specific order or hierarchy described.

[0112] In the above detailed description, various features are grouped together in a single embodiment to simplify the disclosure. This method of disclosure should not be interpreted as reflecting an intention that the embodiments of the claimed subject matter require more features than are clearly stated in each claim. On the contrary, as reflected in the appended claims, the invention is in a state of having less than all the features of the disclosed individual embodiments. Therefore, the appended claims are hereby expressly incorporated into the detailed description, with each claim standing on its own as a separate preferred embodiment of the invention.

[0113] Those skilled in the art will also appreciate that the various illustrative logic blocks, modules, circuits, and algorithmic steps described in conjunction with the embodiments herein can all be implemented as electronic hardware, computer software, or a combination thereof. In order to clearly illustrate the interchangeability between hardware and software, various illustrative components, blocks, modules, circuits, and steps are generally described above around their functions. Whether such functions are implemented as hardware or software depends on specific applications and the design constraints imposed on the entire system. A skilled person can implement the described functions in an alternative manner for each specific application, but such implementation decisions should not be interpreted as departing from the scope of protection of the present disclosure.

[0114] The steps of the methods or algorithms described in connection with the embodiments of the present disclosure may be embodied directly in hardware, in a software module executed by a processor, or in a combination thereof. The software module may be located in a RAM memory, a flash memory, a ROM memory, an EPROM memory, an EEPROM memory, a register, a hard disk, a removable disk, a CD-ROM, or any other form of storage medium well known in the art. An exemplary storage medium is coupled to the processor such that the processor can read information from, and write information to, the storage medium. Of course, the storage medium may also be integral to the processor. The processor and the storage medium may be located in an ASIC. The ASIC may be located in a user terminal. Of course, the processor and the storage medium may also exist as discrete components in the user terminal.

[0115] For a software implementation, the techniques described in this application can be implemented using modules (e.g., procedures, functions, etc.) that perform the functions described in this application. These software codes can be stored in a memory unit and executed by a processor. The memory unit may be implemented within the processor or outside the processor, and in the latter case, it is communicatively coupled to the processor by various means, which are well known in the art.

[0116] Although the embodiments of the present invention have been shown and described, it will be understood by those of ordinary skill in the art that various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principles and spirit of the present invention, and the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A synchronous spatiotemporal multi-step traffic prediction method based on hybrid graph convolution, characterized in that: The following steps are involved: Step 1: According to the periodicity and seasonal characteristics of cellular traffic in time series, the spatiotemporal data set is finely divided into three historical time steps of equal length, and then a gated temporal convolutional network is used to extract the short-term time series features of the traffic; Step 2: According to the randomness and burstiness of cellular traffic in space, the dynamic graph convolutional network based on the spatial domain and the adaptive graph convolutional network based on the spectral domain are integrated to extract the spatial characteristics of the traffic using the hybrid graph structure; Step 3: To address the problem of graph convolutional networks focusing too much on local features, the synchronous spatiotemporal feature fusion module dynamically weights and fuses the spatial features and short-term time series features of the data. After extracting the mixed spatial features, the temporal features carrying historical data are used to dynamically correct this deviation. The formula for the synchronous spatiotemporal feature fusion module is as follows: Among them, W t , W g , b f are all learnable parameters, H l ∈R T×N×D It is the output of the STFM module and also the output of the l-th layer SSTHL module; Step 4: The output of the synchronous spatiotemporal feature fusion module in step 3 is the output of the SSTHL module in the first layer. The parameterized jump connection is introduced to add the output of the previous layer to the input of the current layer. The number of channels of the jump connection and the number of channels for the last time are set. After the jump connection is completed, two fully connected layers are used to realize the output layer. Finally, the multi-step result of the output prediction is: M is the number of time steps to predict into the future.

2. According to claim 1, a synchronous spatiotemporal multi-step traffic prediction method based on hybrid graph convolution is characterized in that: The step 1 comprises the following steps: Step 1.1, the time fine-grained slicing module TSM extracts the cellular traffic spatiotemporal dataset according to different near-time intervals, and divides the fine-grained extraction into three historical time steps of equal length according to near-hour extraction, near-day extraction and near-week extraction; Step 1.2, the fine-grained temporal feature extraction module TFM applies the gate concept and consists of three aggregated gated temporal convolutional networks. Each gated component consists of two dilated causal convolutional blocks to capture the temporal trend of the node, and then uses a multi-layer perceptron MLP to aggregate the three temporal features.

3. According to claim 1, a synchronous spatiotemporal multi-step traffic prediction method based on hybrid graph convolution is characterized in that: The step 2 comprises the following steps: Step 2.1, the spatial domain-based dynamic graph convolutional network ADGL consists of a dynamic graph GAT and a multi-head attention mechanism as well as a priori graph structure that integrates geographic information and semantic information, which is used to effectively extract the dynamic neighborhood of users; Step 2.2, the spectral domain-based adaptive graph convolutional network SAGL is composed of an adaptive adjacency matrix generated based on a random node embedding method or a singular value decomposition method, which is used to effectively extract the bursty random dependency relationship between users and network scheduling; Step 2.3: The spatial correlation of cellular traffic is associated with two graph convolution calculations. Traffic has both stable dynamic spatial neighbor characteristics and hidden random distance spatial dependencies. The calculation results of these two graph structures are weighted and fused.

4. According to claim 1, a synchronous spatiotemporal multi-step traffic prediction method based on hybrid graph convolution is characterized in that: In the step three, different cellular base stations will be affected by different time characteristics and spatial characteristics, and the nodes have different time and space trends. After extracting the mixed spatial characteristics, the time characteristics carrying historical data are used to dynamically correct the deviation, and the gating unit is used to fuse the time series characteristics and spatial characteristics of the traffic.

5. According to claim 4, a synchronous spatiotemporal multi-step traffic prediction method based on hybrid graph convolution is characterized in that: In the step three, four SSTHL modules are stacked, parameterized skip connections are introduced, the output of the previous layer is added to the input of the current layer, and finally two fully connected layers are used to output the final multi-step prediction result.

6. According to claim 2, a synchronous spatiotemporal multi-step traffic prediction method based on hybrid graph convolution is characterized in that: The cellular traffic values ​​extracted at the adjacent time steps are the same. In order to predict the cellular traffic data of M time steps after the future time t, the cellular traffic data of the nearest hour before the time t, that is, T steps, is selected as input. The expression is as follows: X h =(X t-T+1 ,X t-T+2 ,..,X t ),t>T Among them, X h The cellular traffic data is divided into the data set at time t to obtain the data set at time t.

7. The method for synchronous spatiotemporal multi-step traffic prediction based on hybrid graph convolution according to claim 6 is characterized in that: In the cellular traffic with a daily cycle extracted on the near day, the traffic value at a fixed time of the previous day is the same as the traffic value at a fixed time of the next day. In order to predict the cellular traffic data of M time steps after the future time t, the cellular traffic data at the same time of the day before the time t is selected as the starting point of the slice, and the cellular traffic data of T steps is cut back as the input. The expression is as follows: X d =(X t-24×q+1 ,X t-24×q+2 ,..,X t-24×q+T ),t>24×q Among them, X d The cellular traffic data is divided into the adjacent days to obtain the data set at time t.

8. The method for synchronous spatiotemporal multi-step traffic prediction based on hybrid graph convolution according to claim 7 is characterized in that: The extraction of weekly cellular traffic in the near week ignores the difference between the daily rules of weekdays and holidays. Therefore, in the weekly cellular traffic, the traffic value at a fixed time in the previous week is the same as the traffic value at a fixed time in the next week. In order to predict the cellular traffic data of M time steps after the future time t, the cellular traffic data of the week before the time t is selected as the starting point of the slice, and the sliding input of T time steps is cut. The expression is as follows: X w =(X t-24×7×q+1 ,X t-24×7×q+2 ,..,X t-24×7×q+T ),t>7×24×q Among them, X w The cellular traffic data is divided into the adjacent weeks to obtain the data set at time t.

9. The method for synchronous spatiotemporal multi-step traffic prediction based on hybrid graph convolution according to claim 8 is characterized in that: Divide the cellular traffic data into X at time t according to the time, day, and week. h , X d and X w ∈R T×N×F , where N represents the number of nodes and F represents the feature dimension on the node. As time t moves backward, we get a sliding window data set X with different historical information. H , X D , X W , which is expressed as follows: L=(t-7×24)q-M+1 Where L is the length of the sliding window data set.

10. The method for synchronous spatiotemporal multi-step traffic prediction based on hybrid graph convolution according to claim 9, characterized in that: Set X H , X D and X W ∈R L×T×N×F They are input into the local time series feature extraction module LTFM respectively. At this time, it is a multi-node spatiotemporal traffic prediction problem, and only the cellular traffic characteristic values ​​of multiple nodes are predicted. At this time, N>1 and F=1.

Images