An output flow prediction method, device and storage medium
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- CHINA MOBILE COMM LTD RES INST
- Filing Date
- 2022-03-22
- Publication Date
- 2026-07-03
Smart Images
Figure CN116827809B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of communication technology, and in particular to an output flow prediction method, apparatus and storage medium. Background Technology
[0002] Existing machine learning methods include ARIMA (Autoregressive Integrated Moving Average) and SVR (Support Vector Machine Regression). ARIMA predicts future traffic by calculating the average of historical sequences, while SVR is a variation of SVM (Support Vector Machine), which uses machine learning algorithms to fit the relationship between input and output to predict future traffic. Both methods require significant storage space and computing resources.
[0003] In deep learning models, considering the complex spatiotemporal dependencies of base stations, RNNs (Recurrent Neural Networks) are used to solve complex sequence temporal dependencies, CNNs (Convolutional Neural Networks) are used to extract spatial relationships in regular regions, and GCNs (Graph Convolutional Networks) establish irregular topologies for base stations. For example, T-GCN (Time Graph Convolutional Network) combines GRUs (Gated Recurrent Units) and GCNs. GRUs are used to learn the dynamic changes in traffic data to capture temporal dependencies, while GCNs are used to learn complex topologies to capture spatial dependencies. AGG (Anti-Grain Geometry) uses GCNs to capture spatial features of traffic prediction by analyzing the spatial distance between base stations, while TCNs (Time Convolutional Networks) capture temporal dependencies. Furthermore, existing work on base station traffic overload prediction captures the spatiotemporal relationships of the original traffic sequence and directly compares the predicted results with a predefined threshold to determine whether a traffic surge will occur.
[0004] The shortcoming of existing technology is that it learns too much redundant information. Summary of the Invention
[0005] This invention provides an output flow prediction method, apparatus, and storage medium to solve the problem of needing to learn too much redundant information during the prediction process.
[0006] This invention provides the following technical solutions:
[0007] An output flow prediction method, comprising:
[0008] Obtain historical traffic sequences and geographical location information of the base station;
[0009] Identify at least one neighboring base station that is closest in time and space as the neighbor subgraph of the target base station;
[0010] The original traffic sequence and the discrete traffic sequence are obtained by processing the data using threshold setting rules.
[0011] The graph attention network is used to obtain the spatial dependencies between each base station and the target base station in the neighbor subgraph that are similar or opposite in terms of traffic.
[0012] The time dependencies of the original flow sequence and the discrete flow sequence are obtained through TCN;
[0013] The flow prediction results are output based on spatial and temporal dependencies using a soft attention mechanism.
[0014] In practice, the at least one neighboring base station with the closest temporal and spatial dimensions is selected as the neighbor subgraph of the target base station using the Dual-KNN algorithm.
[0015] In practice, determining at least one neighboring base station that is most similar in time and space as the neighbor subgraph of the target base station involves using the KNN algorithm in both the time and space dimensions to determine the neighboring base stations most similar to the target base station. Based on the determined distance and similarity, the nearest preset number of neighboring base stations are selected to form the neighbor subgraph. Cosine similarity is used to determine the time distance, and Euclidean distance is used to determine the spatial distance.
[0016] In practice, when determining spatial dependencies, the weight of the influence of neighboring base stations is determined through GAT.
[0017] In implementation, the weight of the influence of neighboring base stations is determined through GAT, including:
[0018] Linear transformations are used to convert the original sequences of each node to a high-dimensional space, with the base station as the node.
[0019] The pairwise attention coefficients are calculated to represent the importance of node j to node i; positive and negative traffic features are obtained to characterize the similarity or opposite relationship between the target base station and other base stations, respectively.
[0020] The attention scores are normalized using the softmax and softmin functions to obtain the weights of different neighboring base stations, which represent similar and opposite spatial dependencies, respectively.
[0021] Two representations of the base station are obtained by calculating neighbor combinations with different weights, which are calculated from positive and negative relationships respectively;
[0022] GAT output includes its own historical patterns and spatial characteristics, where the spatial characteristics reflect the positive or negative impact of neighboring base stations.
[0023] In implementation, a soft attention mechanism is used to output flow prediction results based on spatial and temporal dependencies, including:
[0024] Extract trend information from the original sequence and overload features from the discrete sequence, and use soft attention to capture the spatiotemporal features of base station traffic;
[0025] Determine the continuous hidden state h of base station i c and discrete hidden state h d Learnable weights are used to calculate the weights and then use a decoder to predict future traffic overload.
[0026] During implementation, it further includes:
[0027] The balancing parameter β is used to alleviate the imbalance in the ratio of positive to negative samples when calculating the error, and the balancing parameter is linearly related to the ratio of positive to negative samples.
[0028] An output flow prediction device, comprising:
[0029] The processor is used to read programs from memory and execute the following procedures:
[0030] Obtain historical traffic sequences and geographical location information of the base station;
[0031] Identify at least one neighboring base station that is closest in time and space as the neighbor subgraph of the target base station;
[0032] The original traffic sequence and the discrete traffic sequence are obtained by processing the data using threshold setting rules.
[0033] The graph attention network is used to obtain the spatial dependencies between each base station and the target base station in the neighbor subgraph that are similar or opposite in terms of traffic.
[0034] The time dependencies of the original flow sequence and the discrete flow sequence are obtained through TCN;
[0035] The flow prediction results are output based on spatial and temporal dependencies using a soft attention mechanism.
[0036] A transceiver is used to receive and send data under the control of a processor.
[0037] In practice, the at least one neighboring base station with the closest temporal and spatial dimensions is selected as the neighbor subgraph of the target base station using the Dual-KNN algorithm.
[0038] In practice, determining at least one neighboring base station that is most similar in time and space as the neighbor subgraph of the target base station involves using the KNN algorithm in both the time and space dimensions to determine the neighboring base stations most similar to the target base station. Based on the determined distance and similarity, the nearest preset number of neighboring base stations are selected to form the neighbor subgraph. Cosine similarity is used to determine the time distance, and Euclidean distance is used to determine the spatial distance.
[0039] In practice, when determining spatial dependencies, the weight of the influence of neighboring base stations is determined through GAT.
[0040] In implementation, the weight of the influence of neighboring base stations is determined through GAT, including:
[0041] Linear transformations are used to convert the original sequences of each node to a high-dimensional space, with the base station as the node.
[0042] The pairwise attention coefficients are calculated to represent the importance of node j to node i; positive and negative traffic features are obtained to characterize the similarity or opposite relationship between the target base station and other base stations, respectively.
[0043] The attention scores are normalized using the softmax and softmin functions to obtain the weights of different neighboring base stations, which represent similar and opposite spatial dependencies, respectively.
[0044] Two representations of the base station are obtained by calculating neighbor combinations with different weights, which are calculated from positive and negative relationships respectively;
[0045] GAT output includes its own historical patterns and spatial characteristics, where the spatial characteristics reflect the positive or negative impact of neighboring base stations.
[0046] In implementation, a soft attention mechanism is used to output flow prediction results based on spatial and temporal dependencies, including:
[0047] Extract trend information from the original sequence and overload features from the discrete sequence, and use soft attention to capture the spatiotemporal features of base station traffic;
[0048] Determine the continuous hidden state h of base station i c and discrete hidden state h d Learnable weights are used to calculate the weights and then use a decoder to predict future traffic overload.
[0049] During implementation, it further includes:
[0050] The balancing parameter β is used to alleviate the imbalance in the ratio of positive to negative samples when calculating the error, and the balancing parameter is linearly related to the ratio of positive to negative samples.
[0051] An output flow prediction device, comprising:
[0052] The data acquisition module is used to acquire historical traffic sequences and geographical location information of the base station;
[0053] The neighbor subgraph module is used to determine at least one neighboring base station that is most similar in time and space as the neighbor subgraph of the target base station;
[0054] The sequence processing module is used to process and obtain the original traffic sequence and the discrete traffic sequence by setting threshold rules;
[0055] The spatial processing module is used to obtain the spatial dependencies between each base station and the target base station in the neighbor subgraph that are similar or opposite in terms of traffic through the graph attention network;
[0056] The time processing module is used to obtain the time dependency relationship between the original flow sequence and the discrete flow sequence through TCN;
[0057] The prediction module is used to output flow prediction results based on spatial and temporal dependencies using a soft attention mechanism.
[0058] In practice, the neighbor subgraph module is further used to select neighbor base stations using the Dual-KNN algorithm when determining at least one neighbor base station with the closest temporal and spatial dimensions as the neighbor subgraph of the target base station.
[0059] In implementation, the neighbor subgraph module is further used to determine the at least one neighbor base station that is most similar in time and space as the neighbor subgraph of the target base station. This is done by using the KNN algorithm in both the time and space dimensions to determine the neighbor base station that is most similar to the target base station. Based on the determined distance and similarity, the nearest preset number of neighbor base stations are selected to form the neighbor subgraph. Cosine similarity is used to determine the time distance, and Euclidean distance is used to determine the spatial distance.
[0060] In practice, the spatial processing module is further used to determine the weight of the influence of neighboring base stations when determining spatial dependencies, by using GAT.
[0061] In implementation, the spatial processing module is further used to determine the weights of neighboring base station influences via GAT, including:
[0062] Linear transformations are used to convert the original sequences of each node to a high-dimensional space, with the base station as the node.
[0063] The pairwise attention coefficients are calculated to represent the importance of node j to node i; positive and negative traffic features are obtained to characterize the similarity or opposite relationship between the target base station and other base stations, respectively.
[0064] The attention scores are normalized using the softmax and softmin functions to obtain the weights of different neighboring base stations, which represent similar and opposite spatial dependencies, respectively.
[0065] Two representations of the base station are obtained by calculating neighbor combinations with different weights, which are calculated from positive and negative relationships respectively;
[0066] GAT output includes its own historical patterns and spatial characteristics, where the spatial characteristics reflect the positive or negative impact of neighboring base stations.
[0067] In implementation, the prediction module is further used to output flow prediction results based on spatial and temporal dependencies using a soft attention mechanism, including:
[0068] Extract trend information from the original sequence and overload features from the discrete sequence, and use soft attention to capture the spatiotemporal features of base station traffic;
[0069] Determine the continuous hidden state h of base station i c and discrete hidden state h d Learnable weights are used to calculate the weights and then use a decoder to predict future traffic overload.
[0070] In practice, the prediction module is further used to alleviate the imbalance in the ratio of positive to negative samples when calculating errors, and the balance parameter is linearly related to the ratio of positive to negative samples.
[0071] A computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described output flow prediction method.
[0072] The beneficial effects of this invention are as follows:
[0073] In the technical solutions provided in the embodiments of this invention, most existing technologies mainly capture the dynamic spatiotemporal dependencies of the original traffic sequence and compare the predicted results with a pre-set threshold to determine whether the traffic is overloaded. However, these solutions contain a large amount of fine-grained redundant information when capturing the temporal features of base stations. Since the discrete sequence after threshold processing contains the required dynamic change pattern of traffic overload, using soft attention to capture the spatiotemporal features of base station traffic can simultaneously extract the trend information in the original sequence and the overload features of the discrete sequence, avoiding the learning of too much redundant information.
[0074] The soft attention mechanism was used to model both the original traffic sequence and the historical sequence of traffic overload, simultaneously capturing trend information from the original sequence and overload features from the discrete sequence. Because the soft attention mechanism can capture features from multiple sequences simultaneously—the original traffic sequence containing the trend of traffic changes, and the thresholded discrete sequence containing the dynamic change pattern of traffic overload—it improves the accuracy of network traffic overload prediction by simultaneously modeling the trend information from the original sequence and the overload features from the discrete sequence.
[0075] In the spatial extraction section, a graph attention network (GAT) is introduced to capture the influence of different weights of neighbors, making the extraction of spatial features more effective. Due to the complex spatial dependencies between base stations caused by human movement, GAT is used to extract spatial dependencies by capturing the spatial influence of different neighbors. Compared with existing methods, it can model the different spatial interactions of neighbors on the target base station, uncovering deeper spatial dependencies.
[0076] Using graph convolutional networks to capture temporal features, TCN (Transformative Convolutional Network) allows for large-scale parallel processing, enabling a larger receptive field and capturing dependencies over longer periods. Because TCN achieves large-scale parallel processing and a larger receptive field through network deepening, it overcomes the low efficiency and gradient vanishing / exploding problems of RNNs, resulting in significant improvements in multi-step prediction accuracy and computational speed. Attached Figure Description
[0077] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this invention, illustrate exemplary embodiments of the invention and are used to explain the invention, but do not constitute an undue limitation of the invention. In the drawings:
[0078] Figure 1 This is a schematic diagram illustrating the implementation process of the output flow prediction method in an embodiment of the present invention;
[0079] Figure 2 This is a schematic diagram of the flow overload prediction system in an embodiment of the present invention;
[0080] Figure 3 This is a schematic diagram of a simple TCN process with a kernel size of 2 in an embodiment of the present invention;
[0081] Figure 4 This is a schematic diagram of the overall structure of the Gated TCN layer in an embodiment of the present invention;
[0082] Figure 5 This is a schematic diagram of the output flow prediction device in an embodiment of the present invention. Detailed Implementation
[0083] The inventor noticed the following during the invention process:
[0084] One approach involves using a random forest algorithm to effectively select features from the original data. Then, a clustering algorithm is used to compare the similarity of the selected features to find the most correlated features. An autoencoder is used to calculate the root mean square error between the original and reduced-dimensional data. Finally, the predicted class label of the sample is obtained by integrating the outputs of each decision tree. Another approach involves inputting time-series network traffic training data samples into an LSTM (Long Short-Term Memory) network for training, obtaining an LSTM-based time-series network traffic prediction model. Test data is then input into the prediction model to obtain predicted network traffic values. Yet another approach uses a ResNet model to capture spatial features and a GRU to capture temporal features. These two types of features are then fused and input into a multi-head attention layer. An attention mechanism is used to assign larger weights to important features, and softmax is used to determine whether network traffic is abnormal. A third approach preprocesses the collected network data, filters out data that meets experimental requirements, converts the processed data into corresponding grayscale images, uses a CNN to capture spatial features and an LSTM model to capture temporal features, determines the optimal parameters of the model by minimizing cross-entropy, and evaluates the detection results based on the trained model.
[0085] The existing similar neural network method is the AGG model, which constructs the topology by calculating the spatial distance between base stations and uses GCN to capture spatial dependencies. However, the spatial dependencies between base stations are quite complex. For example, two base stations with similar functions but far apart may have similar traffic distribution patterns. Therefore, it is particularly important to capture deeper spatial dependencies.
[0086] Some solutions use the random forest algorithm, which is prone to overfitting under high noise conditions, affecting the prediction effect. Other solutions and T-GCN use variants of the RNN model to capture time-dependent features. RNNs need to iteratively train time series, which takes a long time and affects the prediction efficiency. Some solutions use the CNN method to capture the spatial features of the base station, which requires preprocessing the raw data into a spatial grid distribution, resulting in unnecessary calculation errors and affecting the prediction accuracy.
[0087] To address the problems of existing methods, this invention proposes a traffic overload prediction scheme based on temporal graph attention convolutional networks. It uses a Dual-KNN (K-Nearest Neighbor) mechanism to select neighboring base stations to construct a subgraph, and employs GAT to capture spatial dependencies, ensuring the effectiveness of spatial feature extraction. The gated TCN mechanism is used to capture temporal dependencies, solving problems such as long training time and gradient vanishing / exploding in RNN models. Finally, a soft attention mechanism is used to simultaneously capture features of both the original traffic sequence and the thresholded discrete sequence, overcoming the previous drawback of learning excessive redundant information.
[0088] The solution combines TCN, GAT, and soft attention mechanisms to achieve network traffic overload prediction. TCN overcomes the problems of long training time and gradient explosion in RNN models, greatly improving computational efficiency; GAT captures the influence of different weights of neighbors at the spatial level; the soft attention mechanism simultaneously models the original traffic change pattern and the traffic overload occurrence pattern, thereby improving the accuracy of network traffic overload prediction.
[0089] The specific embodiments of the present invention will now be described with reference to the accompanying drawings.
[0090] Figure 1 The flowchart of the output flow prediction method is shown in the figure, and may include:
[0091] Step 101: Obtain the historical traffic sequence of the base station and the geographical location information of the base station;
[0092] Step 102: Determine at least one neighboring base station that is closest in time and space as the neighbor subgraph of the target base station;
[0093] Step 103: Obtain the original traffic sequence and the discrete traffic sequence by processing the data according to the threshold setting rules;
[0094] Step 104: Obtain the spatial dependencies between each base station and the target base station in the neighbor subgraph that are similar or opposite in terms of traffic through the graph attention network;
[0095] Step 105: Obtain the time dependency relationship between the original flow sequence and the discrete flow sequence using TCN;
[0096] Step 106: Output the flow prediction results based on spatial and temporal dependencies using a soft attention mechanism.
[0097] Specifically, the solution aims to predict base station traffic overload by modeling historical traffic patterns of base stations using a combination of TCN (Temporal Convolutional Network) and GAT (Graph Attention Network). A dual-KNN mechanism is used to select neighbors from both spatial and temporal dimensions to construct a subgraph, and GAT is used to capture spatial dependencies, while gated TCN is used to capture temporal dependencies. Considering that soft attention can integrate different features, it is used to simultaneously capture features from both the original traffic sequence and the thresholded discrete sequence.
[0098] The overall system flow will be explained first, followed by an explanation of the implementation of each step.
[0099] Figure 2 This is a flowchart illustrating the traffic overload prediction system. The traffic overload prediction scheme is as follows: Figure 2 As shown, the historical traffic sequence of the input base station and the geographical location information of the base station are used to construct a subgraph by calculating the k base stations that are closest in time and space as neighbors of the target base station. After processing with threshold setting rules, the original traffic sequence and the new discrete traffic sequence are obtained. The time dependency relationship between the original traffic sequence and the discrete traffic sequence is captured by the TCN module. The two hidden state vectors are input into the soft attention mechanism and the output traffic overload prediction result can be obtained by deencoder.
[0100] 1. Implementation of the Dual-KNN mechanism.
[0101] In practice, the at least one neighboring base station with the closest temporal and spatial dimensions is selected as the neighbor subgraph of the target base station using the Dual-KNN algorithm.
[0102] In specific implementation, determining at least one neighboring base station that is most similar in time and space as the neighbor subgraph of the target base station involves using the KNN algorithm in both the time and space dimensions to determine the neighboring base stations most similar to the target base station. Based on the determined distance and similarity, the nearest preset number of neighboring base stations are selected to form the neighbor subgraph. Cosine similarity is used to determine the time distance, and Euclidean distance is used to determine the spatial distance.
[0103] Specifically, in real-world applications, the correlation between many base stations is relatively weak, having little impact on base station traffic surges. To capture the most relevant neighboring base stations to the target base station and extract their spatial features, a Dual-KNN module can be used to select N base stations to construct a subgraph. Since KNN is commonly used to capture the most similar samples of a test target, it can be used in both the temporal and spatial dimensions to capture the most similar neighbors to the target base station. Cosine similarity can be used to capture temporal distance, and Euclidean distance can be used to capture spatial distance. The formulas are as follows:
[0104]
[0105]
[0106] For DIS (Digital Information System), many physical quantities in physics, such as distance, displacement, force, velocity, temperature, pressure, voltage, and current, can be measured using DIS. `dis` represents spatial distance, `lat / lon` represents latitude / longitude, and `sim` represents cosine similarity. i,in ={V i,1 ,…,V i,t} and V n,in ={V n,1 ,…,V n,t |n=1,2,…,N&n≠i} represents the original traffic sequences of the target base station and other base stations, loc i =(lat i ,lon i Let be the geographical location of the i-th base station. Using the above formula, the N nearest base stations can be selected based on the calculated distance and similarity to form a subgraph.
[0107] 2. Graph Attention Network.
[0108] In practice, when determining spatial dependencies, the weight of the influence of neighboring base stations is determined through GAT.
[0109] In practice, the weight of the influence of neighboring base stations is determined through GAT, including:
[0110] Linear transformations are used to convert the original sequences of each node to a high-dimensional space, with the base station as the node.
[0111] The pairwise attention coefficients are calculated to represent the importance of node j to node i; positive and negative traffic features are obtained to characterize the similarity or opposite relationship between the target base station and other base stations, respectively.
[0112] The attention scores are normalized using the softmax and softmin functions to obtain the weights of different neighboring base stations, which represent similar and opposite spatial dependencies, respectively.
[0113] Two representations of the base station are obtained by calculating neighbor combinations with different weights, which are calculated from positive and negative relationships respectively;
[0114] GAT output includes its own historical patterns and spatial characteristics, where the spatial characteristics reflect the positive or negative impact of neighboring base stations.
[0115] Specifically, in order to capture the spatial dependency features of different neighbors in a subgraph, GAT can be used to capture the influence of different neighbor weights.
[0116] First, a linear transformation is used to convert the original sequences of all nodes to a high-dimensional space to improve expressive power. Then, the pairwise attention coefficients e are calculated. i,j This represents the importance of node j to node i. a(·) is a single-layer feedforward neural network, and its formula can be shown below:
[0117]
[0118] Considering the different influences between base stations, it is possible to simultaneously capture positive and negative traffic features, representing the similarity or opposite relationships between the target base station and other base stations, respectively. For example, commuters typically leave residential areas in the morning, resulting in a significant decrease in traffic in residential areas and a rapid increase in traffic in commercial areas. Therefore, the attention scores can be normalized using softmax and softmin functions respectively to obtain the weights of different neighboring base stations, representing similar and opposite spatial dependencies, as shown in the following formula:
[0119]
[0120]
[0121] Exp (exponential function) is an exponential function with the natural constant e as its base.
[0122] Therefore, by calculating neighbor combinations with different weights, two representations of the base station are obtained, which are calculated from positive and negative relationships respectively, as shown in the following formula:
[0123]
[0124] σ(sigma) is used to describe the distribution or dispersion of the average value of any process parameter; σ(n) represents the sum of all positive factors of the integer n.
[0125] Therefore, the final output of the GAT module consists of three parts, including its own historical mode. and spatial features ( and Furthermore, the spatial characteristics reflect the positive and negative impacts of neighboring base stations.
[0126] 3. Gated temporal convolutional layer.
[0127] In implementation, a soft attention mechanism is used to output flow prediction results based on spatial and temporal dependencies, including:
[0128] Extract trend information from the original sequence and overload features from the discrete sequence, and use soft attention to capture the spatiotemporal features of base station traffic;
[0129] Determine the continuous hidden state h of base station i c and discrete hidden state h d Learnable weights are used to calculate the weights and then use a decoder to predict future traffic overload.
[0130] Specifically, Figure 3 This is a simplified TCN process with a kernel size of 2. Considering the flexibility and model performance of temporal convolutional networks, TCN can be applied to extract the time-dependent features of traffic. The simplified process of TCN is as follows: Figure 3 As shown. Compared with RNN, the receptive field of TCN increases exponentially with the number of layers, and it can capture longer temporal dependencies.
[0131] If the kernel size is set to k, then the receptive field size of the l-th layer is (k-1)·l+1. Furthermore, TCN can effectively prevent network overfitting. In summary, assuming the input to the l+1 layer of TCN is h... (l+1) The convolution calculation operation of TCN is as follows:
[0132]
[0133] Among them, h (0) This is the output of the GAT module, W f (l) and b (l) These are learnable parameters, and * indicates a convolution operation. Gated TCN is an improvement on TCN, allowing control over the percentage of information flowing out.
[0134] h (l+1) =f(W f (l) *h (l) +b (l) )⊙σ(W g (l) *h (l) +c(l) )
[0135] Most existing technologies primarily capture the dynamic spatiotemporal dependencies of the original traffic sequence, comparing the predicted results with pre-set thresholds to determine whether traffic is overloaded. However, these schemes contain a large amount of fine-grained redundant information when capturing the temporal features of base stations, while the discrete sequence after thresholding contains the desired dynamic change patterns of traffic overload. Therefore, in order to simultaneously extract trend information from the original sequence and overload features from the discrete sequence, softattention can be used to capture the spatiotemporal features of base station traffic.
[0136] Figure 4 This is a schematic diagram of the overall structure of the Gated TCN layer, as shown below. Figure 4 As shown, a i,c and a i,d The continuous hidden states h of base station i are respectively represented by c and discrete hidden state h d The learnable weights, after weight calculation, can be used by a decoder to predict future traffic overload. The formula is as follows:
[0137]
[0138] MLP stands for Multilayer Perceptron.
[0139] For the predicted overload sequence and the true value sequence Y i,out ={Y i,t+1 ,…,Y i,t+T The error function is calculated using binary cross-entropy, as shown in the formula below:
[0140]
[0141] Loss is the loss function.
[0142] In specific implementation, it may further include:
[0143] The balancing parameter β is used to alleviate the imbalance in the ratio of positive to negative samples when calculating the error, and the balancing parameter is linearly related to the ratio of positive to negative samples.
[0144] Specifically, in error calculation, a balance parameter β can be used to alleviate the imbalance in the ratio of positive to negative samples, and the balance parameter has a linear relationship with the ratio of positive to negative samples. The calculation formula is shown below:
[0145]
[0146] Based on the same inventive concept, this invention also provides an output flow prediction device and a computer-readable storage medium. Since the principle of these devices in solving the problem is similar to that of the output flow prediction method, the implementation of these devices can refer to the implementation of the method, and the repeated parts will not be described again.
[0147] When implementing the technical solutions provided in the embodiments of the present invention, they can be implemented in the following manner.
[0148] Figure 5 The figure shows a schematic diagram of the output flow prediction device. The device includes:
[0149] Processor 500 is used to read the program from memory 520 and execute the following procedures:
[0150] Obtain historical traffic sequences and geographical location information of the base station;
[0151] Identify at least one neighboring base station that is closest in time and space as the neighbor subgraph of the target base station;
[0152] The original traffic sequence and the discrete traffic sequence are obtained by processing the data using threshold setting rules.
[0153] The graph attention network is used to obtain the spatial dependencies between each base station and the target base station in the neighbor subgraph that are similar or opposite in terms of traffic.
[0154] The time dependencies of the original flow sequence and the discrete flow sequence are obtained through TCN;
[0155] The flow prediction results are output based on spatial and temporal dependencies using a soft attention mechanism.
[0156] Transceiver 510 is used to receive and send data under the control of processor 500.
[0157] In practice, the at least one neighboring base station with the closest temporal and spatial dimensions is selected as the neighbor subgraph of the target base station using the Dual-KNN algorithm.
[0158] In practice, determining at least one neighboring base station that is most similar in time and space as the neighbor subgraph of the target base station involves using the KNN algorithm in both the time and space dimensions to determine the neighboring base stations most similar to the target base station. Based on the determined distance and similarity, the nearest preset number of neighboring base stations are selected to form the neighbor subgraph. Cosine similarity is used to determine the time distance, and Euclidean distance is used to determine the spatial distance.
[0159] In practice, when determining spatial dependencies, the weight of the influence of neighboring base stations is determined through GAT.
[0160] In implementation, the weight of the influence of neighboring base stations is determined through GAT, including:
[0161] Linear transformations are used to convert the original sequences of each node to a high-dimensional space, with the base station as the node.
[0162] The pairwise attention coefficients are calculated to represent the importance of node j to node i; positive and negative traffic features are obtained to characterize the similarity or opposite relationship between the target base station and other base stations, respectively.
[0163] The attention scores are normalized using the softmax and softmin functions to obtain the weights of different neighboring base stations, which represent similar and opposite spatial dependencies, respectively.
[0164] Two representations of the base station are obtained by calculating neighbor combinations with different weights, which are calculated from positive and negative relationships respectively;
[0165] GAT output includes its own historical patterns and spatial characteristics, where the spatial characteristics reflect the positive or negative impact of neighboring base stations.
[0166] In implementation, a soft attention mechanism is used to output flow prediction results based on spatial and temporal dependencies, including:
[0167] Extract trend information from the original sequence and overload features from the discrete sequence, and use soft attention to capture the spatiotemporal features of base station traffic;
[0168] Determine the continuous hidden state h of base station i c and discrete hidden state h d Learnable weights are used to calculate the weights and then use a decoder to predict future traffic overload.
[0169] During implementation, it further includes:
[0170] The balancing parameter β is used to alleviate the imbalance in the ratio of positive to negative samples when calculating the error, and the balancing parameter is linearly related to the ratio of positive to negative samples.
[0171] Among them, Figure 5In this context, the bus architecture may include any number of interconnected buses and bridges, specifically linking various circuits together, represented by one or more processors (processor 500) and memory (memory 520). The bus architecture may also link various other circuits such as peripheral devices, voltage regulators, and power management circuits, which are well known in the art and therefore will not be described further herein. The bus interface provides an interface. The transceiver 510 may be multiple elements, including transmitters and receivers, providing a unit for communicating with various other devices over a transmission medium. The processor 500 is responsible for managing the bus architecture and general processing, and the memory 520 may store data used by the processor 500 during operation.
[0172] This invention also provides an output flow prediction device, comprising:
[0173] The data acquisition module is used to acquire historical traffic sequences and geographical location information of the base station;
[0174] The neighbor subgraph module is used to determine at least one neighboring base station that is most similar in time and space as the neighbor subgraph of the target base station;
[0175] The sequence processing module is used to process and obtain the original traffic sequence and the discrete traffic sequence by setting threshold rules;
[0176] The spatial processing module is used to obtain the spatial dependencies between each base station and the target base station in the neighbor subgraph that are similar or opposite in terms of traffic through the graph attention network;
[0177] The time processing module is used to obtain the time dependency relationship between the original flow sequence and the discrete flow sequence through TCN;
[0178] The prediction module is used to output flow prediction results based on spatial and temporal dependencies using a soft attention mechanism.
[0179] In practice, the neighbor subgraph module is further used to select neighbor base stations using the Dual-KNN algorithm when determining at least one neighbor base station with the closest temporal and spatial dimensions as the neighbor subgraph of the target base station.
[0180] In implementation, the neighbor subgraph module is further used to determine the at least one neighbor base station that is most similar in time and space as the neighbor subgraph of the target base station. This is done by using the KNN algorithm in both the time and space dimensions to determine the neighbor base station that is most similar to the target base station. Based on the determined distance and similarity, the nearest preset number of neighbor base stations are selected to form the neighbor subgraph. Cosine similarity is used to determine the time distance, and Euclidean distance is used to determine the spatial distance.
[0181] In practice, the spatial processing module is further used to determine the weight of the influence of neighboring base stations when determining spatial dependencies, by using GAT.
[0182] In implementation, the spatial processing module is further used to determine the weights of neighboring base station influences via GAT, including:
[0183] Linear transformations are used to convert the original sequences of each node to a high-dimensional space, with the base station as the node.
[0184] The pairwise attention coefficients are calculated to represent the importance of node j to node i; positive and negative traffic features are obtained to characterize the similarity or opposite relationship between the target base station and other base stations, respectively.
[0185] The attention scores are normalized using the softmax and softmin functions to obtain the weights of different neighboring base stations, which represent similar and opposite spatial dependencies, respectively.
[0186] Two representations of the base station are obtained by calculating neighbor combinations with different weights, which are calculated from positive and negative relationships respectively;
[0187] GAT output includes its own historical patterns and spatial characteristics, where the spatial characteristics reflect the positive or negative impact of neighboring base stations.
[0188] In implementation, the prediction module is further used to output flow prediction results based on spatial and temporal dependencies using a soft attention mechanism, including:
[0189] Extract trend information from the original sequence and overload features from the discrete sequence, and use soft attention to capture the spatiotemporal features of base station traffic;
[0190] Determine the continuous hidden state h of base station i c and discrete hidden state h d Learnable weights are used to calculate the weights and then use a decoder to predict future traffic overload.
[0191] In practice, the prediction module is further used to alleviate the imbalance in the ratio of positive to negative samples when calculating errors, and the balance parameter is linearly related to the ratio of positive to negative samples.
[0192] For ease of description, the various parts of the device described above are divided into modules or units according to their functions. Of course, in implementing this invention, the functions of each module or unit can be implemented in one or more software or hardware components.
[0193] This invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described output flow prediction method.
[0194] For details on implementation, please refer to the implementation of the output flow prediction method.
[0195] In summary, the technical solution provided in this invention uses a soft attention mechanism to model both the original traffic sequence and the historical sequence of traffic overload, simultaneously capturing trend information in the original sequence and overload characteristics of the discrete sequence. Since the soft attention mechanism can capture features of multiple sequences simultaneously—the original traffic sequence containing the trend of traffic changes, and the thresholded discrete sequence containing the dynamic change pattern of traffic overload—the use of the soft attention mechanism to simultaneously model both the trend information in the original sequence and the overload characteristics of the discrete sequence improves the accuracy of network traffic overload prediction.
[0196] In the spatial extraction section, a graph attention network (GAT) is introduced to capture the influence of different weights of neighbors, making the extraction of spatial features more effective. Due to the complex spatial dependencies between base stations caused by human movement, GAT is used to extract spatial dependencies by capturing the spatial influence of different neighbors. Compared with existing methods, it can model the different spatial interactions of neighbors on the target base station, uncovering deeper spatial dependencies.
[0197] Using graph convolutional networks to capture temporal features, TCN (Transformative Convolutional Network) allows for large-scale parallel processing, enabling a larger receptive field and capturing dependencies over longer periods. Because TCN achieves large-scale parallel processing and a larger receptive field through network deepening, it overcomes the low efficiency and gradient vanishing / exploding problems of RNNs, resulting in significant improvements in multi-step prediction accuracy and computational speed.
[0198] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage and optical storage) containing computer-usable program code.
[0199] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.
[0200] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.
[0201] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.
[0202] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.
Claims
1. An output flow prediction method characterized by, include: Obtain historical traffic sequences and geographical location information of the base station; Identify at least one neighboring base station that is closest in time and space as the neighbor subgraph of the target base station; The original traffic sequence and the discrete traffic sequence are obtained by processing the data using threshold setting rules. The spatial dependencies between each base station and the target base station in the neighbor subgraph are obtained by using a graph attention network to identify similar or opposite traffic relationships. The temporal dependencies between the original flow sequence and the discrete flow sequence are obtained by passing them through a temporal convolutional network (TCN). The flow prediction results are output based on spatial and temporal dependencies using a soft attention mechanism. Among them, determining at least one neighboring base station with the closest temporal and spatial dimensions as the neighbor subgraph of the target base station is achieved by selecting neighboring base stations using the Dual-KNN algorithm.
2. The method of claim 1, wherein, Determining at least one neighboring base station that is most similar to the target base station in both time and space is achieved by using the K-Nearest Neighbors (KNN) algorithm in both the time and space dimensions to identify the neighboring base stations most similar to the target base station. Based on the determined distance and similarity, a predetermined number of the nearest neighboring base stations are selected to form the neighboring subgraph. Cosine similarity is used to determine the time distance, and Euclidean distance is used to determine the spatial distance.
3. The method of claim 1, wherein, When determining spatial dependencies, the weights of the influence of neighboring base stations are determined through a graph attention network (GAT).
4. The method of claim 3, wherein, The weights of neighboring base station influence are determined using GAT, including: Linear transformations are used to convert the original sequences of each node to a high-dimensional space, with the base station as the node. The pairwise attention coefficients are calculated to represent the importance of nodes to each other; positive and negative traffic features are obtained to characterize the similarity or opposite relationship between the target base station and other base stations, respectively. The attention scores are normalized using the softmax and softmin functions to obtain the weights of different neighboring base stations, which represent similar and opposite spatial dependencies, respectively. Two representations of the base station are obtained by calculating neighbor combinations with different weights, which are calculated from positive and negative relationships respectively; GAT output includes its own historical patterns and spatial characteristics, where the spatial characteristics reflect the positive or negative impact of neighboring base stations.
5. The method of claim 1, wherein, The flow prediction results are output based on spatial and temporal dependencies using a soft attention mechanism, including: Extract trend information from the original sequence and overload features from the discrete sequence, and use soft attention to capture the spatiotemporal features of base station traffic; The learnable weights of the continuous and discrete hidden states of the base station are determined, and the future traffic overload is predicted by the decoder after the weights are calculated.
6. The method of claim 5, wherein, Further includes: A balancing parameter is used to alleviate the imbalance in the ratio of positive to negative samples when calculating the error, and the balancing parameter has a linear relationship with the ratio of positive to negative samples.
7. An output flow prediction device, characterized by, include: The processor is used to read programs from memory and execute the following procedures: Obtain historical traffic sequences and geographical location information of the base station; The target base station is determined by selecting at least one neighboring base station that is most similar in time and space as its neighbor subgraph. This selection is achieved through the Dual-KNN algorithm. The original traffic sequence and the discrete traffic sequence are obtained by processing the data using threshold setting rules. The spatial dependencies between each base station and the target base station in the neighbor subgraph are obtained by using a graph attention network to identify similar or opposite traffic relationships. The time dependencies of the original flow sequence and the discrete flow sequence are obtained through TCN; The flow prediction results are output based on spatial and temporal dependencies using a soft attention mechanism. A transceiver is used to receive and send data under the control of a processor.
8. An output flow prediction device, characterized in that, include: The data acquisition module is used to acquire historical traffic sequences and geographical location information of the base station; The neighbor subgraph module is used to determine at least one neighbor base station that is most similar in time and space as the neighbor subgraph of the target base station. The determination of at least one neighbor base station that is most similar in time and space as the neighbor subgraph of the target base station is carried out by selecting neighbor base stations through the Dual-KNN algorithm. The sequence processing module is used to process and obtain the original traffic sequence and the discrete traffic sequence by setting threshold rules; The spatial processing module is used to obtain the spatial dependencies between each base station and the target base station in the neighbor subgraph that are similar or opposite in terms of traffic through the graph attention network; The time processing module is used to obtain the time dependency relationship between the original flow sequence and the discrete flow sequence through TCN; The prediction module is used to output flow prediction results based on spatial and temporal dependencies using a soft attention mechanism.
9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the method of any one of claims 1 to 6.