Tree convolutional network model method and system for traffic flow prediction

Abstract

The invention discloses a tree convolutional network model method and system for traffic flow prediction, and the method comprises the steps: receiving traffic data, and abstracting a traffic map structure; abstracting a traffic map structure into nodes and a communication relationship thereof to obtain a spatial relationship of node distribution; initializing rear nodes and constructing a space tree matrix; performing a tree convolution operation on the spatial tree matrix, and performing feature aggregation on a convolution result to generate a high-dimensional traffic flow feature representation; and constructing a prediction model based on a deep learning network, and inputting the high-dimensional traffic flow features into the prediction model to output a traffic state prediction result of the target time step. By constructing the space tree matrix, the space-time dependency relationship of the traffic flow can be more accurately captured, so that the space relevance and the time sequence change between the traffic nodes are effectively combined, the modeling capability of the complex relationship between the traffic nodes is enhanced, and finally the space tree matrix is processed through the tree convolution operation. And generating high-dimensional traffic flow feature representation.

Description

Technical Field

[0001] The present invention relates to the technical field of traffic state prediction, and in particular to a tree convolutional network model method and system for traffic flow prediction. Background Art

[0002] With the development of intelligent transportation systems, traffic flow prediction has become a vital part of urban traffic management. Traditional traffic flow prediction methods are mainly based on statistical models and classical machine learning methods, such as time series analysis, regression analysis, support vector machines, etc. These methods have good results in some scenarios. However, these methods have the following limitations: (1) Unable to fully capture spatiotemporal dependencies: When processing time series data, traditional methods usually rely on direct inference of historical data and ignore the spatiotemporal dependencies between characteristics such as traffic flow, speed and occupancy. In particular, when there is significant spatial correlation between traffic nodes, these methods are difficult to effectively model and predict. (2) Poor model generalization ability: Traditional models are usually trained based on specific assumptions. When the traffic network topology changes or emergencies occur, the prediction effect of traditional models is prone to deviation and lacks good generalization ability. (3) Difficulty in processing high-dimensional data: With the deployment of a large number of sensors and monitoring equipment in intelligent transportation systems, the amount and dimension of collected data have increased dramatically. When processing these high-dimensional data, traditional methods have low computational efficiency and it is difficult to extract effective information from large amounts of data. In addition, the mainstream methods of traffic flow prediction, such as the traffic flow prediction model based on graph convolutional network (GCN), although they can handle certain spatial correlations, still have the following problems: (1) Graph convolutional networks are difficult to capture complex spatiotemporal dependencies: When dealing with spatial correlations in traffic networks, graph convolutional networks can effectively capture the connection relationship between nodes, but they often cannot provide sufficient information when modeling temporal dynamics and long-term series dependencies. (2) Lack of effective feature fusion mechanism: Existing models have limitations in the fusion of spatiotemporal features and cannot effectively combine the temporal dynamic characteristics and spatial distribution characteristics of traffic flow, resulting in insufficient prediction accuracy of the model. (3) Unable to cope with complex traffic flow change patterns: Traffic flow changes are often affected by multiple factors, such as traffic accidents, weather changes, road construction, etc., and these changes often have sudden and nonlinear characteristics. In summary, traditional traffic flow prediction methods and existing graph convolutional networks are difficult to quickly adapt to these changes. Summary of the invention

[0003] In view of this, the present invention proposes a tree convolution network model method and system for traffic flow prediction, which can extract the spatiotemporal characteristics of traffic flow by constructing a spatial tree matrix and applying tree convolution operations, thereby realizing the prediction of traffic flow. The present invention provides the following technical solutions:

[0004] A tree convolutional network model method for traffic flow prediction, the method comprising: receiving traffic data, abstracting a traffic graph structure based on the traffic data; abstracting nodes and their connectivity relationships based on the traffic graph structure to obtain a preliminary spatial relationship of node distribution; constructing a spatial tree matrix based on initialized nodes; performing a tree convolution operation on the spatial tree matrix, and performing feature aggregation on the convolution result to generate a high-dimensional traffic flow feature representation; constructing a prediction model based on a deep learning network, and inputting the high-dimensional traffic flow features into the prediction model to output a traffic flow prediction result for a target time step.

[0005] Optionally, the method for receiving traffic data and abstracting a traffic graph structure based on the traffic data includes: collecting real-time traffic data from different locations in the traffic network, the traffic data including at least traffic flow, traffic speed and traffic occupancy rate; abstracting the collection location of the traffic data as a node in a graph structure, and storing the traffic data collected by all nodes as time series data; defining the connectivity relationship between different collection locations according to the road topology structure of the actual traffic network, abstracting the connectivity relationship as the edge of the graph structure, and calculating the edge weight; generating an adjacency matrix according to the edge weights between nodes; and generating a traffic graph structure according to the nodes, edges, adjacency matrix and time series data.

[0006] Optionally, the method of abstracting nodes and their connectivity relationships based on the traffic graph structure to obtain preliminary spatial relationships of node distribution includes: extracting features of each node from the nodes of the traffic graph structure and storing them as a multidimensional feature vector, including time series features and spatial position features; classifying the spatial distribution types of nodes through the adjacency matrix in the traffic graph structure, including random uniform distribution and small-scale node cluster distribution; identifying cluster centers, and assigning nodes to corresponding cluster areas based on the distance or connectivity between the nodes and the cluster centers, and then initializing the hierarchical relationships of the nodes to obtain preliminary spatial relationships of node distribution.

[0007] Optionally, the method for constructing a spatial tree matrix based on initialized nodes includes: taking each node in the initialized node distribution as a root node; traversing the nodes layer by layer starting from the root node through breadth-first search to generate multiple plane tree structures; converting the multiple plane tree structures into corresponding plane tree matrices; and superimposing and fusing the multiple plane tree matrices into a spatial tree matrix in the order of the root nodes.

[0008] Optionally, the method of performing a tree convolution operation on the spatial tree matrix and performing feature aggregation on the convolution results to generate a high-dimensional traffic flow feature representation includes: defining a tree convolution kernel based on the spatial tree matrix, and the convolution kernel slides along the hierarchical direction of the tree structure and the node connection direction; performing hierarchical convolution on each layer of the spatial tree matrix to capture the associated features of nodes at different levels; performing directional convolution on each path of the spatial tree matrix to extract directional features from leaf nodes to root nodes; performing feature aggregation on the results of hierarchical convolution and directional convolution, and fusing node features using weighted summation or maximum pooling methods; and outputting a high-dimensional traffic flow feature representation for input to subsequent prediction models.

[0009] Optionally, the method of constructing a prediction model based on a deep learning network and inputting high-dimensional traffic flow features into the prediction model to output traffic flow prediction results for a target time step includes: constructing a traffic flow prediction model based on a deep learning network and inputting the high-dimensional traffic flow features into the traffic flow prediction model; extracting the temporal dynamics and spatial distribution characteristics of the high-dimensional traffic flow features through the traffic flow prediction model; calculating the output value of the prediction model, the output value including the traffic flow, average speed and road occupancy rate at the target time step; performing error evaluation on the output value and optimizing the prediction model based on real historical data; and outputting the traffic flow prediction results for the target time step based on the optimized prediction model.

[0010] The present invention further discloses a tree convolutional network model system for traffic flow prediction, comprising:

[0011] A graph structure building module, used for receiving traffic data and abstracting a traffic graph structure based on the traffic data;

[0012] A node analysis module is used to abstract nodes and their connectivity relationships based on the traffic graph structure and obtain a preliminary spatial relationship of node distribution;

[0013] A tree matrix construction module is used to construct a spatial tree matrix based on the initialized nodes and their connectivity relationships;

[0014] A tree convolution module, used for performing a tree convolution operation on the spatial tree matrix and performing feature aggregation on the convolution result to generate a high-dimensional traffic flow feature representation;

[0015] The state prediction module is used to build a prediction model based on the deep learning network and input high-dimensional traffic flow features into the prediction model to output the traffic flow prediction results at the target time step.

[0016] The present invention further discloses a computer-readable storage medium, which stores a computer program. When the computer program is executed by a processor, the tree convolutional network model method for traffic flow prediction is implemented.

[0017] The present invention further discloses an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the above-mentioned tree convolutional network model method for traffic flow prediction when executing the program.

[0018] The present invention further discloses a computer program product, including a computer program, which implements the above-mentioned tree convolutional network model method for traffic flow prediction when executed by a processor.

[0019] According to the technical solution of the present invention, by constructing a spatial tree matrix based on the traffic graph structure, the spatiotemporal dependency of traffic flow can be captured more accurately, thereby effectively combining the spatial correlation and time series changes between traffic nodes, and improving the accuracy of traffic flow prediction. At the same time, the construction of the spatial tree matrix further enhances the modeling ability of complex relationships between traffic nodes, effectively overcoming the problem that traditional methods are difficult to accurately model spatial relationships in complex networks, and finally achieving the processing of the spatial tree matrix through tree convolution operations to generate high-dimensional traffic flow feature representation. Through the feature aggregation method, key information of traffic flow can be extracted from multiple dimensions, including flow, speed, occupancy, etc., providing richer input features for subsequent traffic status prediction. This high-dimensional feature representation can significantly improve the performance of the prediction model in the face of a changing traffic environment and reduce prediction errors. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] For the purpose of illustration and not limitation, the present invention is now described in conjunction with the embodiments of the present invention and the accompanying drawings, in which:

[0021] Figure 1 It is a schematic diagram of the flow of a tree convolutional network model method for traffic flow prediction in an embodiment of the present invention;

[0022] Figure 2 It is a schematic diagram of the structure of a tree convolutional network model system for traffic flow prediction in an embodiment of the present invention;

[0023] Figure 3 is a schematic structural diagram of an electronic device in an embodiment of the present invention;

[0024] Figure 4 is a schematic diagram of traffic flow prediction values ​​at different time steps in an embodiment of the present invention;

[0025] Figure 5 is a schematic diagram of comparative evaluation results of traffic occupancy rate and average speed in a random uniform distribution scenario in an embodiment of the present invention;

[0026] Figure 6It is a schematic diagram of the effect of the number of nodes on the model in an embodiment of the present invention;

[0027] Figure 7 is another schematic diagram of the effect of the number of nodes on the model in an embodiment of the present invention;

[0028] Figure 8 It is a schematic diagram of statistical results of three indicators in a small-scale node aggregation scenario in an implementation manner of the present invention;

[0029] Fig. 9 is a schematic diagram of evaluation results of traffic occupancy rate and average speed in a small-scale node aggregation scenario in an implementation manner of the present invention;

[0030] Fig.10 It is a schematic diagram of predicted values ​​and true values ​​of the tree convolution method and other baselines in the random uniform distribution and small-scale node aggregation scenarios in the implementation mode of the present invention;

[0031] Fig.11 It is a schematic diagram of the comparison results of the root mean square error in the random uniform distribution and small-scale node aggregation scenarios in the implementation mode of the present invention;

[0032] Fig.12 It is a schematic diagram of predicted values ​​and actual values ​​of traffic flows output by different models in an embodiment of the present invention;

[0033] Fig.13 is a schematic diagram of predicted values ​​and true values ​​of average traffic speed output by different models in an embodiment of the present invention;

[0034] Fig.14 is a residual graph of traffic speed prediction values ​​and true values ​​output by different models in the implementation mode of the present invention;

[0035] Fig.15 Schematic diagram of prediction results of the tree convolutional network for different periods in an embodiment of the present invention;

[0036] Fig.16 is a schematic diagram of a spatial tree matrix design process in an embodiment of the present invention;

[0037] Fig.17 Schematic diagram of the process of capturing the directional features of the plane tree convolution in an embodiment of the present invention;

[0038] Fig.18 It is a schematic diagram of the process of capturing the plane tree convolutional hierarchical features in an embodiment of the present invention. DETAILED DESCRIPTION

[0039] In order to enable those skilled in the art to better understand the solution of the present application, the technical solution in the implementation mode of the present application will be clearly and completely described below in conjunction with the drawings in the implementation mode of the present application. Obviously, the described implementation mode is only a part of the implementation mode of the present application, not all the implementation modes. Based on the implementation mode in the present application, all other implementation modes obtained by ordinary technicians in the field without creative work should fall within the scope of protection of the present application.

[0040] In addition, some of the above terms may be used to express other meanings in addition to indicating orientation or positional relationship. For example, the term "on" may also be used to express a certain dependency or connection relationship in some cases. For those of ordinary skill in the art, the specific meanings of these terms in this application can be understood according to the specific circumstances. In addition, the term "plurality" should mean two or more than two.

[0041] It should be noted that, in the absence of conflict, the embodiments of the present application and the features in the embodiments can be combined with each other. The embodiments of the present application will be described in detail below with reference to the accompanying drawings.

[0042] refer to Figure 1 This embodiment discloses a tree convolutional network model method for traffic flow prediction, taking the aggregation of traffic variables (such as traffic flow, traffic speed and traffic occupancy rate) as an example to predict traffic flow. The method includes:

[0043] S100: Receive traffic data, and abstract a traffic graph structure based on the traffic data, that is, traffic time series data in a graph structure scenario. The traffic data in this embodiment are all from real monitoring data, which are obtained by monitoring sensors installed on the highway. The data collected by each sensor includes the traffic status within a specific time period, and the data is organized into a time series according to timestamps. In the traffic graph structure of this embodiment, the monitoring sensors are nodes, and the edges between different nodes are defined according to the actual road topology of the traffic network. The weight of the edge can be calculated based on factors such as road distance, road capacity or flow correlation, and finally generate an adjacency matrix of the traffic graph. The adjacency matrix is ​​used to represent the connection relationship between nodes. Finally, a traffic graph structure is generated based on nodes, edges, adjacency matrix and time series data. In a real-time application scenario, real-time traffic data at different locations in the traffic network are collected, and the traffic data at least includes traffic flow, traffic speed and traffic occupancy rate; the collection location of the traffic data is abstracted as a node in the graph structure, and the traffic data collected by all nodes is stored as time series data; the connectivity relationship between different collection locations is defined according to the road topology of the actual traffic network, and the connectivity relationship is abstracted as an edge of the graph structure, and the edge weight is calculated. Specifically, traffic data includes traffic flow, traffic speed and traffic occupancy rate. Traffic flow refers to the number of vehicles detected by a single sensor per unit time. Traffic speed refers to the average speed of all vehicles detected by the sensor per unit time, in kilometers per hour. Traffic occupancy rate refers to the proportion of traffic roads occupied by vehicles per unit time. The above are all exemplary data.

[0044] S200: Abstract the nodes and their connectivity relationships based on the traffic graph structure to obtain the preliminary spatial relationship of the node distribution. Extract time series features and spatial position features from each node in the traffic graph structure constructed in step S100. The time series features can be traffic flow, traffic speed and traffic occupancy rate, and the spatial position features can be the geographic coordinates or relative positions corresponding to the nodes. Based on this, the features of each node are stored as a multidimensional feature vector, which includes time series data and spatial location information. Further, the spatial distribution of the nodes is classified according to the adjacency matrix in step S100 to identify nodes with different distributions, including random uniform distribution and small-scale node cluster distribution. Then the hierarchical relationship of the nodes is initialized.

[0045] S300: Construct a spatial tree matrix based on the initialized nodes and their connectivity relationships. After obtaining the distribution of all nodes, select each node in the traffic graph structure as the root node, and construct the tree structure around the idea of ​​data regularization. First, abstract the nodes and their spatial relationships according to the graph structure to achieve the preliminary spatial relationship of the node distribution. Secondly, construct a plane tree matrix with different nodes as the root nodes of the tree. Finally, the plane tree matrices of all nodes are fused into a spatial tree matrix representing the global spatial relationship of the nodes. In order to further overcome the problem of unbalanced correlation in small-scale node aggregation scenarios, this implementation method fully considers the spatial connectivity between nodes. It adds a hierarchical design and directed information transfer design to the tree structure, which helps to use deep models to implement time series data mining and prediction tasks. The dimension of the plane tree matrix can reflect the hierarchical and directional characteristics of the tree structure. The hierarchical feature is reflected in the fact that nodes at the same level are distributed in the same row of the tree matrix. The directional feature is reflected in the ability to accurately track and construct all nodes connected in a directional manner according to the order of each column of the plane tree matrix. The spatial tree matrix is ​​a three-dimensional matrix with spatial characteristics, and its design process is as follows Fig.16 As shown. Each node is set as the root node, and the traceability of all nodes is completed using breadth-first traversal, thus forming different tree structures. The three-dimensional space tree matrix is ​​composed of the plane tree matrix formed by all tree structures. Fig.16 The conversion result from the idealized graph structure to the tree structure is shown, which means that the number of layers of the tree structure generated by the breadth-first method for each node is the same. However, the node distribution in the actual application scenario is complex, which makes the generated tree structure level complex and changeable. In order to overcome this problem, the present invention defines the dimension of each plane tree matrix of the spatial tree matrix as Among them, α represents the maximum number of layers of all tree structures. α-1 is the maximum number of nodes at the bottom layer. Multiple plane tree matrices are superimposed and merged into a spatial tree matrix in the order of root nodes to generate the final spatial tree matrix.

[0046] S400: performing a tree convolution operation on the spatial tree matrix, and performing feature aggregation on the convolution result to generate a high-dimensional traffic flow feature representation.

[0047] Specifically, based on the spatial tree matrix, a tree convolution kernel is defined, and the convolution kernel slides along the hierarchical direction of the tree structure and the node connection direction. The spatial tree convolution operation is completed by combining multiple plane tree convolution operations. The main task of the plane tree convolution is to construct a multi-layer convolution structure oriented to the plane tree matrix, so that the process can make full use of the spatial correlation of the nodes in the tree structure. The edges connecting the nodes of the tree structure contain information with weights. In order to compress the weight information of the edges into the same value range, the present invention uses the standard deviation normalization method to normalize the edge data. The combined use of the tree matrix and the plane tree convolution can effectively capture the directional features between nodes, such as Fig.17 As shown. Each column of the plane tree matrix stores the connection path from the leaf node to the root node. Each node's connection path has an independent convolution process, which will not cross the connection paths of other nodes. The connection features of each node's connection path are converted into high-level features. Finally, the tree convolution connects the high-level features of all connection paths into the direction feature of the root node.

[0048] Fig.18 The process of capturing hierarchical features in tree convolution is shown. Since the tree structure is generated by breadth-first traversal of the graph, the tree structure and the plane tree matrix can store the hierarchical relationship between nodes. The hierarchical relationship is interpreted as a direct or indirect relationship connected to the current node. The root node is always stored in the first row of the plane tree matrix. The nodes directly connected to the root node are stored in the second row of the plane tree matrix. Other nodes indirectly connected to the root node are stored in different rows of the plane tree matrix according to the number of indirect connections. The plane tree convolution obtains the hierarchical relationship of each row of the plane tree matrix by the translation calculation of the convolution kernel according to the row priority principle. The node layer with a large number of indirect connections is first convolved, which means that the indirectly connected nodes are subjected to more convolution processing. Each convolution process must have a certain degree of information loss. Therefore, the root node and the directly connected nodes can retain more directly related information because they perform fewer convolution processes. The above process enables the high-level features generated by the tree convolution to contain different degrees of hierarchical features between nodes.

[0049] Hierarchical convolution is performed on each layer of the spatial tree matrix to capture the associated features of nodes at different levels; directional convolution is performed on each path of the spatial tree matrix to extract the directional features from the leaf node to the root node. For the hierarchical direction, feature extraction is performed layer by layer from the leaf node to the root node along the hierarchical relationship of the tree structure to capture the associated features between nodes at different levels. Specifically, for each layer of nodes in the spatial tree matrix, the tree convolution kernel slides in the hierarchical direction to weightedly fuse the features of the current node with the features of the parent node; the information of the current node is transmitted to the parent node through the tree convolution operation, and the child node feeds back the information of the parent node through the aggregation operation, thereby updating the node features. For the node connection direction, along the path between the nodes, the information transmission and associated features between the parent node and the child node are captured, that is, the directional features from the leaf node to the root node are extracted to reflect the directional association between the nodes. Specifically, the convolution kernel slides along the path direction of the tree structure, and convolution is performed from the leaf node upward in sequence to extract the feature changes on the path; the directional information of each path is retained and fused into the feature representation of the root node step by step through the convolution operation.

[0050] The results of hierarchical convolution and directional convolution are subjected to feature aggregation, and the node features are fused using weighted summation or maximum pooling methods to output a high-dimensional traffic flow feature representation for input into subsequent prediction models. This embodiment exemplifies the following aggregation methods: (1) Weighted summation: The features of each node are weighted and accumulated according to the weight to highlight the key node features; (2) Maximum pooling: The largest eigenvalue is selected in each convolution window to retain the most significant local information; (3) Mean pooling: The node features are averaged within the convolution window to smooth the information between nodes. Specifically, the convolution results of each layer are aggregated to obtain the feature representation of the layer, and the features of all layers are aggregated to the root node to generate a high-dimensional feature vector. The node features after feature aggregation are fused into a high-dimensional traffic flow feature representation. This high-dimensional feature representation can comprehensively include the following information: (1) local hierarchical features of nodes: the hierarchical relationship and local correlation of nodes in the tree; (2) directional features of paths: the direction of information transmission from leaf nodes to root nodes; (3) global spatial distribution features: the spatial position relationship and correlation between different nodes; (4) temporal dynamic features: the historical traffic status change trend reflected in the node time series data.

[0051] S500: Build a prediction model based on the deep learning network, and input high-dimensional traffic flow features into the prediction model to output the traffic flow prediction results at the target time step.

[0052] The prediction model in this embodiment is constructed based on a deep learning network to meet the needs of spatiotemporal dynamic modeling of high-dimensional traffic flow characteristics. The network includes but is not limited to:

[0053] Recurrent Neural Networks (RNN / LSTM / GRU): used to model the temporal dynamic characteristics of traffic flow and suitable for capturing short-term and long-term dependencies in time series;

[0054] Convolutional Neural Network (CNN): used to further extract local spatial patterns in high-dimensional traffic flow features;

[0055] Graph Neural Network (GNN): Suitable for capturing complex spatial correlations between nodes and further optimizing the spatial feature modeling of transportation networks.

[0056] Network layer design:

[0057] Input layer: receives the high-dimensional traffic flow feature representation generated from step S400;

[0058] Hidden layer: stacking multiple layers of networks (such as LSTM layers, convolutional layers, or graph convolutional layers) to achieve spatiotemporal joint modeling;

[0059] Output layer: Generates the traffic status prediction result at the target time step through the fully connected layer or regression output layer.

[0060] Specifically, the high-dimensional traffic flow features generated in step S400 are input into the prediction model, and the historical traffic data is used as training data. The model learns the mapping relationship between the features and the future traffic state. The output of the prediction model includes: (1) traffic flow: the expected number of vehicles passing per unit time; (2) average speed: the expected average speed of vehicles per unit time; (3) road occupancy rate: the proportion of roads occupied by vehicles per unit time, ranging from [0,1]. The output format can be a single-step prediction (such as the state of the next 5 minutes) or a multi-step prediction (such as a continuous sequence of states in the next 1 hour).

[0061] After obtaining the model output results, the error of the model output results is evaluated using real historical data. Common evaluation indicators include: (1) Mean square error (MSE): evaluates the square error between the predicted value and the true value; (2) Mean absolute error (MAE): evaluates the average absolute deviation between the predicted value and the true value; (3) Symmetric mean absolute percentage error (SMAPE): suitable for error evaluation when the dynamic range of traffic data is large. Based on the error evaluation results, the model parameters are optimized through the back propagation algorithm, the network weights are updated, and the model is iteratively trained to reduce the prediction error. Finally, based on the optimized deep learning network, the final target time step traffic state prediction result is output. Supports multi-granularity traffic prediction: (1) Local node prediction: predicts traffic, speed or occupancy for a specific node (such as a sensor); (2) Global prediction: integrates the prediction results of all nodes to generate the state distribution of the entire traffic network.

[0062] In the context of time series prediction, this embodiment quantitatively and qualitatively compares the tree convolution method of the present invention with the seasonal autoregressive integrated moving average model, the support vector machine method, the convolutional neural network method, the long short-term memory method, the graph convolutional neural network method, the spatiotemporal graph convolution method based on the attention mechanism, and the spatiotemporal network method based on the multi-head attention mechanism, as follows:

[0063] (1) Quantitative comparison:

[0064] In order to fully reflect the advantages of tree convolution, the embodiments of the present invention quantitatively compare the tree convolution method with other seven existing technologies in random uniform distribution and small-scale node aggregation scenarios. The data used for quantitative comparison includes three categories, namely traffic flow, road occupancy rate and traffic speed.

[0065] A: Quantitative comparison under random uniform distribution scenario

[0066] refer to Figure 4 and Figure 5 , shows the quantitative analysis results of traffic flow prediction for all methods at different time steps in the future. The prediction errors of all methods gradually increase with the increase of time step, which is a basic phenomenon in time series prediction tasks. The spatial structure of the tree convolution method can show good prediction ability for traffic flow data within the next 45 minutes, but the error begins to increase rapidly when predicting traffic flow data for the next 60 minutes. The tree convolution method focuses on the capture of spatial features and lacks a capture module that focuses on temporal features, so it lacks the prediction ability for long-term time series tasks. Reference Figure 6 , unlike the tree convolution method, the effect of the graph convolution neural network increases with the increase of the number of nodes. For example, as the number of nodes increases, the graph convolution method can establish a connection relationship. The source of node feature information enables the model to capture more information about adjacent nodes. Although the convolution neural network does not have the non-Euclidean spatial structure analysis capability similar to the graph convolution neural network, the present invention performs cascade and convolution operations on the time series data of adjacent nodes when designing the convolution neural network. Therefore, the effect of the convolution neural network is also affected by the number of nodes. The long short-term memory neural network and the gated recurrent unit are recursive models that only rely on the time series data of the node itself, so the regularity of its prediction results is only reflected in the time step. The tree convolution method of the present invention has the best prediction effect for 50 nodes, and the number of nodes has a significant effect on the prediction effect of the tree convolution method. When the number of nodes is small, the number of layers of the tree structure composed of nodes is small, and the hierarchical characteristics between nodes cannot be reflected. In addition, when the number of nodes is large, the sparsity of the tree structure increases, resulting in an increase in the number of sparse values ​​in the spatial tree matrix.

[0067] The tree convolution method of the present invention has significant improvements on the prediction results of the three types of traffic data, and the improvement on the prediction effect of traffic occupancy rate and average speed is more obvious. Figure 5 As shown. The reason for this phenomenon is that both traffic occupancy rate and average speed have a fixed range of values. Traffic occupancy rate, as a proportional data, is a proportional data, and its value range is [0,1]. The average speed represents the average speed of all vehicles within a certain time range, and its range of variation is affected by the driver's driving habits and road speed limits. However, traffic flow represents the total number of vehicles on the road, which is the cumulative sum of the data. Therefore, traffic flow data has an unfixed range of variation and a large variation characteristic. The tree convolution method of the present invention and other existing technologies have put forward higher challenges to the prediction ability of traffic flow data, which also makes the research on the tree convolution method more in line with real-life scenarios.

[0068] B: Quantitative comparison in small-scale node cluster distribution scenarios

[0069] refer to Figure 7 ,The effect of the time step on random uniform distribution and small-scale node clustering distribution is the same, that the error increases with the increase of the time step.,However, the effect of the number of nodes on the traffic flow prediction in the small-scale node clustering,distribution scenario is different from the random uniform distribution, e. Figure 7 As shown in the figure. Although tree convolution and graph convolution have almost the same effect in the small-scale node cluster distribution scenario of 20 nodes, compared with graph convolution, the model effect of tree convolution in the small-scale node cluster distribution scenario of 50 nodes and 80 nodes is significantly improved.

[0070] refer to Figure 8 and Fig. 9 , showing the results of quantitative comparison of different methods in small-scale clustered distribution scenarios. Taking the tree convolution traffic flow comparison results as an example, the root mean square error is reduced by 45.2% on average, the mean absolute error is reduced by 42.98% on average, and the symmetric mean absolute percentage error is reduced by 40.05% on average. In the small-scale node clustering scenario, as the number of nodes increases, the small-scale clustering characteristics in the region are enhanced. The differences in the changes in time series data between different regions are also gradually obvious. The graph convolutional neural network is a model that acts on the global graph, and it does not capture the process of local hierarchical features. Similarly, the prediction results of convolutional neural networks, long short-term memory neural networks, and gated recurrent units also show that their effects are flawed compared with the tree convolution method.

[0071] (2) Qualitative comparison:

[0072] refer to Fig.10 , Fig.10(a) and (b) are the prediction results of the tree convolution method and convolutional neural network in the random uniform distribution and small-scale node aggregation scenarios. Convolutional neural networks play an important role in the application of neural networks. A single-layer convolutional neural network can capture local short-term feature-related information through the smooth movement of the convolution kernel. The advantage of convolutional neural networks is that the captured features are very rich and easy to train. However, the disadvantage of convolutional neural networks is that they cannot capture extensive contextual information in the prediction task of time series data. This is because a single-layer convolutional neural network cannot capture long-distance features, which is limited by the receptive field of the convolution kernel. Fig.10 The comparison results shown in (a) and (b) demonstrate the shortcomings of convolutional neural networks. Whether in random uniform distribution or in small-scale node aggregation scenarios, convolutional neural networks lack the ability to fit local peaks. From a macro perspective, the time series data of traffic flow has a relatively concentrated distribution interval. The inability of convolutional neural networks to capture long-distance features makes the model unable to form gradient weights for long-distance peaks. Fig.10 (c) to (f) show the prediction results of the tree convolution method, gated recurrent unit, and long short-term memory neural network in the random uniform distribution and small-scale node aggregation scenarios. Long short-term memory neural network and gated recurrent unit are recurrent neural networks. Their calculation process is similar to that of convolutional neural network. They produce results by applying forward calculation and update the gradient of the model using back propagation. Convolutional neural network, long short-term memory neural network and gated recurrent unit allow multiple neurons to coexist and allow multiple layers of neural network to be connected to each other. The difference between them is that long short-term memory neural network and gated recurrent unit are used to describe the output of time continuous state. The memory mechanism and gating mechanism of long short-term memory neural network and gated recurrent unit can capture long-term dependencies, which is conducive to solving long-term linear sequence problems. Therefore, the prediction effect of long short-term memory neural network and gated recurrent unit on time series data is better than that of convolutional neural network. However, long short-term memory neural network and gated recurrent unit can only capture the information characteristics of the time dimension, but lack the capture of spatial node correlation characteristics. Therefore, the prediction results of long short-term memory neural network and gated recurrent unit do not match the local peak. The above phenomenon is more obvious in the prediction results of small-scale node aggregation scenario. Convolutional neural networks, long short-term memory neural networks, and gated recurrent units cannot process data with non-Euclidean structures. The data design process of these methods cannot take into account the spatial distribution characteristics of the data. If the time series data is designed as a graph structure, the spatial distribution characteristics of the graph can be captured. The spatial distribution characteristics of the graph include node characteristics (the temporal characteristics of a single node itself) and structural characteristics (the interaction characteristics between all nodes in the graph). The graph convolutional network can automatically learn the node characteristics and structural characteristics of all nodes. Fig.10(g) and (h) show the prediction results of the graph convolutional network in random uniform distribution and small-scale node aggregation scenarios. Each convolutional layer of the graph convolutional network can process the node information of the first-order neighborhood and realize the information transmission of multi-order neighborhoods by stacking the node information of multiple convolutional layers. The above process makes the graph convolutional network work well in random uniform distribution. However, the prediction results of the graph convolutional network show that there is a certain underfitting problem in the local peak in the small-scale node aggregation scenario. The time changes of nodes in a single aggregation area in the small-scale node aggregation scenario are very similar, but the time changes between different areas are very different. The graph convolutional network cannot capture the hierarchical characteristics of the spatial distribution of nodes, and it can only use connectivity for adjacency information transmission. The tree convolutional network not only increases the computational amount of directional features in the adjacency information transmission, but also abstracts the differences between aggregation areas into hierarchical features. More importantly, the tree convolutional network and the graph convolutional network have different representation methods for spatial distribution structures. The calculation process of the graph convolutional network aims to obtain a unique global graph containing the spatial distribution of all nodes. In contrast, each node in the tree convolutional network has a spatially distributed tree structure with each node as a root node, which enables the tree convolutional network to capture deeper and more complex spatial correlation features between nodes.

[0073] (3) Comparison between traditional time series methods and space-time models:

[0074] refer to Fig.11 As a classic time series forecasting method, the seasonal difference autoregressive moving average model can capture the periodicity and trend characteristics of traffic data. However, the seasonal difference autoregressive moving average model exhibits weak nonlinear modeling capabilities. In traffic data, there are usually complex nonlinear characteristics, which may hinder the seasonal difference autoregressive moving average model from accurately capturing the dynamic changes in complex traffic scenarios. Fig.12 Traffic flow forecast and Fig.13 The average speed prediction result is shown in the figure. In traffic time series prediction, support vector machines are used to establish nonlinear relationships between input variables (such as historical traffic data) and target variables (such as future traffic flow and traffic speed). However, support vector machines are sensitive to noise and outliers, especially in high-dimensional data. This is Fig.12 and Fig.13 It is obvious that the support vector machine has insufficient fitting ability for local peaks.

[0075] Spatiotemporal model is an important research method for traffic flow prediction, which studies traffic data from the spatial and temporal dimensions. Traffic data shows significant correlation in the changes of traffic data at different time and spatial locations, and the spatiotemporal model can integrate these correlations to improve the effectiveness of traffic prediction. The combination of graph convolutional network and temporal model allows the spatiotemporal model to capture spatial correlation characteristics and temporal dependency characteristics. The present invention compares the prediction performance of tree convolutional network, spatiotemporal graph convolution method based on attention mechanism, spatiotemporal network method based on multi-head attention mechanism, and the combination of tree convolutional network and temporal model.

[0076] Compared with the tree convolutional network, the spatiotemporal model has better traffic flow prediction effect. Among them, the two spatiotemporal models based on graph convolutional networks integrate additional time models and add attention mechanisms to make the model structure more complex and better capture time-dependent features. In contrast, the tree convolutional network is similar to the graph convolutional network and is more suitable for capturing the spatial relationship between traffic nodes rather than time-dependent features. Therefore, the prediction ability of the tree convolutional network is relatively weak. However, when the tree convolutional network is combined with the time model, it can achieve similar application effects as the graph convolutional network in the spatiotemporal model. In addition, the tree convolutional network can effectively capture the directional characteristics and hierarchical characteristics of nodes in the traffic network and reduce the prediction error in the traffic spatiotemporal model. The statistical results of the root mean square error are shown in Figure 2. Fig.11 shown. Fig.12 Traffic flow results are shown, showing qualitative comparisons of predicted and actual values ​​within consecutive time steps. The prediction performance of the combination of tree convolutional networks and temporal models is fully demonstrated in two datasets: in random uniform distribution and small-scale node cluster distribution scenarios, the average error is reduced by 36% and 38.7% respectively compared with existing spatiotemporal models. Traffic speed has different data characteristics from traffic flow, and its uniqueness plays a vital role in traffic research and intelligent transportation systems. Traffic speed is a dynamic variable that changes continuously over time and space. Significant changes in traffic speed may occur at different times and places. Traffic speed does not usually show a linear relationship. Under high traffic flow, traffic speed usually tends to decrease as vehicle density increases. For example, Fig.13 The truth curve in shows the temporal variation of the average speed of all vehicles collected by the current node throughout the day. With a time step of 5 minutes, there are 288 time steps in a day, covering 24 hours. The speed value of each time step represents the average speed of all vehicles collected by the current node within a 5-minute interval. The fluctuation of the truth curve shows that in the early morning and at night, there are fewer vehicles passing through the current node, resulting in minimal traffic congestion and a higher average vehicle speed. On the contrary, during the morning peak hours from 7 to 9 a.m. and the afternoon peak hours from 2 to 4 p.m., when the number of vehicles is large, there will be a certain degree of traffic congestion, resulting in a lower average vehicle speed.

[0077] The prediction results of traffic speed by different models are as follows: Fig.13 and Fig.14 shown. Fig.13 The following table shows the prediction results of 6 different models for the continuous change of traffic speed at the current node throughout the natural day. In order to fully illustrate the difference in prediction performance between the 6 models, Fig.14 A residual plot is given. Fig.14 In the example, the horizontal axis represents the range of values ​​predicted by the model, while the vertical axis represents the difference between the model prediction and the actual value. The residual plot provides a clear visualization of the deviation between the model prediction and the actual value, with the difference close to 0 indicating a more accurate prediction performance. Fig.14 , we can clearly observe that the combination of the tree convolutional network and the temporal model produces residual results with a distribution close to 0. After that, the two spatiotemporal models based on graph convolutional networks show similar characteristics, while the residual results of the seasonal difference autoregressive moving average model and the support vector machine are far below 0, indicating relatively poor prediction performance. Although all six models produce fitting curves similar to the actual values, they differ in their ability to fit local peaks. The tree convolutional network and the two spatiotemporal models based on graph convolutional networks consider more complex node spatial structures, enabling them to better capture the complex changes in traffic speed. In contrast, the seasonal difference autoregressive moving average model and the support vector machine are traditional time series methods that can only process relatively simple short-term time information, resulting in poor performance when processing high dynamic speed data. The combination of the tree convolutional network and the temporal model is able to feature fuse the topological structure information of the traffic network with the time-related features of the time series data, further improving the prediction performance of the model.

[0078] (4) Comparison of model generalization ability:

[0079] In order to verify the generalization ability of the tree convolutional network, the present invention uses three different sets of data at different periods to evaluate the prediction performance of the tree convolutional network, such as Fig.15 shown. Fig.15 (a) shows the prediction results of the tree convolutional network using the daily cycle dataset during the model training and testing stages. Fig.15 (b) and (c) show the prediction results of the tree convolutional network using the weekly cycle dataset and the monthly cycle dataset during the model training and testing phases, respectively. Fig.15 The data in (b) and (c) do not participate in the training process of the tree convolutional network.

[0080] Fig.15 The results show that although there is no fitting Fig.15(b) and (c) show local peaks in the data, but the tree convolution network is able to capture the trend of traffic data changes. The tree convolution network uses a mean square error regularization function to reduce model complexity during training. In addition, the tree convolution network only maintains the parameter information of the spatial tree convolution module during the entire training process. This process provides a relatively simple model structure for the tree convolution network, which helps to improve its model generalization ability.

[0081] In summary, by constructing a spatial tree matrix and performing tree convolution operations, the present invention can simultaneously capture the spatial distribution characteristics of nodes in the traffic network and the temporal dynamic characteristics of time series data, and realize spatiotemporal joint modeling. Through tree convolution operations and feature aggregation, the generated high-dimensional traffic flow feature representation is more comprehensive and accurate, which helps the deep learning network to improve the prediction accuracy of complex traffic conditions. By supporting traffic network modeling with random uniform distribution and small-scale node cluster distribution, it adapts to different road topological structures and is particularly suitable for peak congestion and emergency scenarios. Through deep learning networks and feedback optimization mechanisms, this method can continuously update model parameters, adapt to dynamically changing traffic conditions, and improve generalization performance. The output traffic flow prediction results can provide real-time decision support for intelligent traffic management systems, including congestion warning, path optimization, signal light scheduling and other measures to alleviate traffic pressure and improve road resource utilization.

[0082] refer to Figure 2 This embodiment further discloses a tree convolutional network model system for traffic flow prediction, including:

[0083] The graph structure construction module 21 is used to receive traffic data and abstract a traffic graph structure based on the traffic data; it is also used to collect real-time traffic data at different locations in the traffic network, and the traffic data at least includes traffic flow, traffic speed and traffic occupancy rate; the collection location of the traffic data is abstracted as a node in the graph structure, and the traffic data collected by all nodes is stored as time series data; the connectivity relationship between different collection locations is defined according to the road topology structure of the actual traffic network, and the connectivity relationship is abstracted as the edge of the graph structure, and the edge weight is calculated; an adjacency matrix is ​​generated according to the edge weights between nodes; and a traffic graph structure is generated according to the nodes, edges, adjacency matrix and time series data.

[0084] The node analysis module 22 is used to abstract the nodes and their connectivity relationships based on the traffic map structure to obtain the preliminary spatial relationship of the node distribution; it is also used to extract the features of each node from the nodes of the traffic map structure and store them as a multi-dimensional feature vector, including time series features and spatial position features; classify the spatial distribution types of the nodes through the adjacency matrix in the traffic map structure, including random uniform distribution and small-scale node cluster distribution; identify the cluster center, and assign the nodes to the corresponding cluster area according to the distance or connectivity between the nodes and the cluster center, and then initialize the hierarchical relationship of the nodes to obtain the preliminary spatial relationship of the node distribution.

[0085] The tree matrix construction module 23 is used to construct a spatial tree matrix based on the initialized nodes and their connectivity relationships; and is also used to use each node in the initialized node distribution as a root node; and to traverse the nodes layer by layer starting from the root node through breadth-first search to generate multiple plane tree structures. The multiple plane tree structures are converted into corresponding plane tree matrices; and the multiple plane tree matrices are superimposed and merged into a spatial tree matrix in the order of the root nodes.

[0086] The tree convolution module 24 is used to perform a tree convolution operation on the spatial tree matrix and perform feature aggregation on the convolution results to generate a high-dimensional traffic flow feature representation; it is also used to define a tree convolution kernel based on the spatial tree matrix, and the convolution kernel slides along the hierarchical direction of the tree and the node connection direction; perform hierarchical convolution on each layer of the spatial tree matrix to capture the associated features of nodes at different levels; perform directional convolution on each path of the spatial tree matrix to extract directional features from leaf nodes to root nodes; perform feature aggregation on the results of hierarchical convolution and directional convolution, and fuse node features using weighted summation or maximum pooling methods; output a high-dimensional traffic flow feature representation for input to subsequent prediction models.

[0087] The state prediction module 25 is used to build a prediction model based on a deep learning network, and input high-dimensional traffic flow features into the prediction model to output the traffic flow prediction result of the target time step; it is also used to build a traffic flow prediction model based on a deep learning network, and input the high-dimensional traffic flow features into the traffic flow prediction model; extract the time dynamics and spatial distribution characteristics of the high-dimensional traffic flow features through the traffic flow prediction model; calculate the output value of the prediction model, the output value includes the traffic flow, average speed and road occupancy rate of the target time step; perform error evaluation on the output value, and optimize the prediction model based on real historical data; output the traffic flow prediction result of the target time step based on the optimized prediction model.

[0088] Figure 3 A schematic diagram of the physical structure of an electronic device provided in an embodiment of the present invention, such as Figure 3As shown, the electronic device 50 includes: a processor 501 (processor), a memory 502 (memory) and a bus 503;

[0089] The processor 501 and the memory 502 communicate with each other via the bus 503 ; the processor 501 is used to call program instructions in the memory 502 to execute the methods provided by the above-mentioned method implementation methods.

[0090] This embodiment provides a non-transitory computer-readable storage medium, which stores computer instructions. The computer instructions enable a computer to execute the methods provided by the above-mentioned method embodiments.

[0091] A person skilled in the art can understand that all or part of the steps for implementing the above-mentioned method implementation method can be completed by hardware related to program instructions, and the aforementioned program can be stored in a computer-readable storage medium, which, when executed, executes the steps of the above-mentioned method implementation method; and the aforementioned storage medium includes: ROM, RAM, magnetic disk or optical disk, etc., various storage media that can store program codes.

[0092] The device implementation described above is merely illustrative, wherein the units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, i.e., they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the present implementation scheme. Those of ordinary skill in the art may understand and implement it without creative effort.

[0093] Through the description of the above implementation methods, those skilled in the art can clearly understand that each implementation method can be implemented by means of software plus a necessary general hardware platform, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solution is essentially or the part that contributes to the prior art can be embodied in the form of a software product, and the computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a disk, an optical disk, etc., including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods of each implementation method or some parts of the implementation method.

[0094] The above specific implementations do not constitute a limitation on the protection scope of the present invention. It should be understood by those skilled in the art that various modifications, combinations, sub-combinations and substitutions may occur depending on design requirements and other factors. Any modification, equivalent substitution and improvement made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A tree convolutional network model method for traffic flow prediction, characterized in that: The method comprises: receiving traffic data, and abstracting a traffic graph structure based on the traffic data; Abstracting nodes and their connectivity relationships according to the traffic graph structure, and obtaining a preliminary spatial relationship of node distribution; Construct a spatial tree matrix based on the initialized nodes and their connectivity relationships; Performing a tree convolution operation on the spatial tree matrix and performing feature aggregation on the convolution result to generate a high-dimensional traffic flow feature representation; A prediction model is built based on a deep learning network, and high-dimensional traffic flow features are input into the prediction model to output the traffic flow prediction results at the target time step.

2. The tree convolutional network model method for traffic flow prediction according to claim 1 is characterized in that: The method of receiving traffic data and abstracting a traffic graph structure based on the traffic data includes: Collecting real-time traffic data at different locations in the traffic network, wherein the traffic data at least includes traffic flow, traffic speed and traffic occupancy rate; Abstracting the collection location of traffic data as a node in a graph structure, and storing the traffic data collected by all nodes as time series data; Define the connectivity relationship between different collection locations according to the road topology of the actual traffic network, abstract the connectivity relationship into the edge of the graph structure, and calculate the edge weight; Generate an adjacency matrix based on the weights of nodes and edges between them; Generate traffic graph structures from nodes, edges, adjacency matrices, and time series data.

3. The tree convolutional network model method for traffic flow prediction according to claim 1, characterized in that: The method for abstracting nodes and their connectivity relationships based on the traffic graph structure and obtaining a preliminary spatial relationship of node distribution includes: Extracting the features of each node from the nodes of the traffic graph structure and storing them as a multi-dimensional feature vector, including time series features and spatial location features; The spatial distribution types of nodes are classified through the adjacency matrix in the traffic graph structure, including random uniform distribution and small-scale node cluster distribution; The clustering centers are identified, and nodes are assigned to corresponding clustering areas according to their distance or connectivity to the clustering centers. Then, the hierarchical relationship of the nodes is initialized to obtain the preliminary spatial relationship of the node distribution.

4. The tree convolutional network model method for traffic flow prediction according to claim 1, characterized in that: The method for constructing a spatial tree matrix based on initialized nodes comprises: Take each node in the initialized node distribution as the root node; Through breadth-first search, nodes are traversed layer by layer starting from the root node to generate multiple planar tree structures; Converting the plurality of plane tree structures into corresponding plane tree matrices; Multiple plane tree matrices are superimposed and fused into a spatial tree matrix in the order of root nodes.

5. The tree convolutional network model method for traffic flow prediction according to claim 1, characterized in that: The method of performing a tree convolution operation on the spatial tree matrix and performing feature aggregation on the convolution result to generate a high-dimensional traffic flow feature representation includes: Based on the spatial tree matrix, a tree convolution kernel is defined, and the convolution kernel slides along the hierarchical direction and the node connection direction of the tree structure; Perform hierarchical convolution on each layer of the spatial tree matrix to capture the correlation features of nodes at different levels; Perform directional convolution on each path of the spatial tree matrix to extract directional features from leaf nodes to root nodes; Aggregate the results of hierarchical convolution and directional convolution, and fuse node features using weighted summation or maximum pooling methods; Output high-dimensional traffic flow feature representation for input into subsequent prediction models.

6. The tree convolutional network model method for traffic flow prediction according to claim 1, characterized in that: The method of constructing a prediction model based on a deep learning network and inputting high-dimensional traffic flow features into the prediction model to output a traffic flow prediction result at a target time step includes: Building a traffic flow prediction model based on a deep learning network, and inputting the high-dimensional traffic flow features into the traffic flow prediction model; Extracting the temporal dynamics and spatial distribution characteristics of the high-dimensional traffic flow characteristics through the traffic flow prediction model; Calculating output values ​​of the prediction model, the output values ​​including traffic flow, average speed and road occupancy rate at a target time step; Performing error evaluation on the output value and optimizing the prediction model based on real historical data; The traffic flow prediction results at the target time step are output based on the optimized prediction model.

7. A tree convolutional network model system for traffic flow prediction, characterized in that: include: A graph structure building module, used for receiving traffic data and abstracting a traffic graph structure based on the traffic data; A node analysis module is used to abstract nodes and their connectivity relationships based on the traffic graph structure and obtain a preliminary spatial relationship of node distribution; A tree matrix construction module is used to construct a spatial tree matrix based on the initialized nodes and their connectivity relationships; A tree convolution module, used for performing a tree convolution operation on the spatial tree matrix and performing feature aggregation on the convolution result to generate a high-dimensional traffic flow feature representation; The state prediction module is used to build a prediction model based on the deep learning network and input high-dimensional traffic flow features into the prediction model to output the traffic flow prediction results at the target time step.

8. A computer-readable storage medium, characterized in that: The storage medium stores a computer program, and when the computer program is executed by a processor, the method according to any one of claims 1 to 6 is implemented.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, the method described in any one of claims 1 to 6 is implemented.

10. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the method according to any one of claims 1 to 6 is implemented.

Images