Traffic prediction system and method based on fine-grained association mining

Abstract

The traffic flow prediction system based on fine correlation mining comprises the following modules: an input convolution module, a correlation capturing function module, a space-time feature processing module and an output convolution module; the traffic flow prediction method based on fine correlation mining comprises the following operation steps: (1) the input convolution module maps low-dimensional historical traffic flow data into high-dimensional historical traffic flow data; (2) the correlation capturing function module adaptively extracts the correlation between traffic points and calculates n hypergraphs representing the correlation between traffic points; (3) the space-time feature processing module calculates l high-dimensional feature matrices of each traffic point which fuse space-time information; (4) the output convolution module calculates the prediction value of the future traffic point flow speed according to the high-dimensional feature matrices of the traffic points.

Description

TECHNICAL FIELD

[0001] The present application relates to a traffic flow prediction system and method based on fine correlation mining, belongs to the field of information technology, and particularly belongs to the field of intelligent transportation information technology. BACKGROUND

[0002] In recent years, with the rapid development of the Internet of Things, the development of intelligent transportation systems has been increasingly valued in the construction of smart cities. By organically combining sensor technology, data communication technology, and computer technology for the entire transportation management system, intelligent transportation systems help to achieve accurate and efficient transportation management in a wider road network. Traffic flow is a basic indicator reflecting the road condition, which is represented by the traffic speed collected by sensors deployed at road traffic points. It is usually in the form of time series data. Traffic flow prediction is the most important part of intelligent transportation systems and an important prerequisite for implementing intelligent transportation management. Since the road network traffic data has a strong correlation in both time and space dimensions, how to model the correlation of the data and make accurate traffic flow prediction in a complex road network environment is a challenging problem.

[0003] In order to model the complex spatio-temporal correlation of traffic data, the existing technical solutions usually model the traffic flow index data into a graph structure to represent the correlation between traffic flow indexes using the inherent topological structure of the road network, usually in the form of an adjacency matrix. Deep learning techniques are used to mine valuable temporal patterns from the data and aggregate features between indexes with strong correlation, thereby considering the features of both time and space dimensions to make predictions.

[0004] In actual road traffic, the traffic flow of different points of the road network has the following characteristics: first, multiple consecutive points on the same road usually exhibit similar time characteristics in a group, and have a relatively consistent reflection on the changes of the traffic conditions of other points of the road network; different size point sets may exhibit different commonalities, and such commonalities associate the points with each other; second, the traffic flow index data of a point in the road network is usually affected by the traffic conditions of multiple other points, and the association between traffic points is usually multi-element, for example, the impact of a sudden traffic accident at a point on the traffic flow of other points on the same road and points on another road without intersection should be different. The prior art models the road network traffic information as a simple graph structure, taking the traffic points as nodes, and the edges in the graph can only represent the pair-wise association strength of the traffic flow indexes of the two points and cannot cover the commonalities between the point groups. At the same time, the prior art lacks detailed description of the edge attributes in the process of graph modeling, and therefore cannot reflect the multi-element nature of the association between points. The above shortcomings result in the lack of comprehensive consideration of the traffic flow association modeling between points of the road network in the prior art, and further result in inaccurate traffic flow prediction of traffic points.

[0005] In summary, how to effectively solve the accurate prediction of traffic flow in smart cities has become a technical problem urgently needed to be solved in the field of intelligent transportation technology. SUMMARY

[0006] Therefore, the purpose of the present application is to provide a system and method for predicting road network traffic flow based on multi-view hypergraph, which can capture more complex and real association between traffic points, describe the potential types of association between traffic points, aggregate the time and space characteristics of traffic flow indexes, and achieve more accurate traffic flow prediction.

[0007] In order to achieve the above purpose, the present application provides a traffic flow prediction system based on fine association mining, which comprises the following modules:

[0008] Input convolution module: the function of this module is to map the low-dimensional historical traffic flow data sensed by the road traffic point sensor to high-dimensional historical traffic flow data, and then input the high-dimensional historical traffic flow data to the spatio-temporal processing module; the input convolution module is composed of a standard convolution network layer with a convolution kernel size of 1x1;

[0009] Association capturing function module: the function of this module is to adaptively extract the association between traffic points, that is, to calculate n hypergraphs representing the association between traffic points from n different perspectives; the nodes in the hypergraph correspond to traffic points; each hyperedge connects several nodes, representing the shared association between the connected traffic points; the hypergraph is represented by an adjacency matrix where subscript i represents the hypergraph of the ith view, p i represents the size of the ith view, and N represents the number of traffic points deployed on the road; E i represents the set of hyperedges in the hypergraph of the ith view; the correlation capturing function module is composed of a multi-layer perception MLP network connection composed of tanh units and ReLU units;

[0010] The space-time feature processing module is composed of l network structures of the same space-time feature processing layer, and each space-time feature processing layer is composed of a time convolution function sub-module and a correlation modeling function sub-module. The function of the module is to receive the high-dimensional historical traffic flow data output by the input convolution module and the n hypergraphs output by the correlation capturing function module, calculate l high-dimensional feature matrices of each traffic point position fused with space-time information, and then output to the output convolution module.

[0011] The output convolution module is to directly sum the l high-dimensional feature matrices of each traffic point position output by the space-time feature processing module, and then map the high-dimensional feature matrices after the sum to a low-dimensional traffic flow data matrix, i.e. the predicted value of the traffic point position flow speed in the future set time period, by using a 1x1 standard convolution layer.

[0012] The specific process of the correlation capturing function module for calculating n hypergraphs representing the correlation between traffic points from n different views includes the following operation steps:

[0013] (21) Given historical traffic flow data where T p represents the length of the historical period, and N represents the number of traffic points deployed on the road; the correlation capturing function module first randomly initializes a spatial feature matrix for the N traffic points where D1 is the set spatial feature dimension, and then generates a time feature for the N traffic points according to the timestamp information of the collected data, i.e. a time feature matrix where D2 is the set time feature dimension; by concatenating the spatial feature matrix and the time feature matrix in the feature dimension, and using a multi-layer perception MLP to encode the concatenated matrix, a space-time feature matrix of the traffic point position fused with space-time information is obtained. where each row of the space-time feature matrix E represents the space-time feature vector of the corresponding numbered traffic point position.

[0014] (22) According to step (21), two spatial feature matrices E s1 and E s2, respectively, and the time feature matrix E t The two matrices are spliced, and then a multi-layer perception is used to encode the two spliced matrices to obtain two spacial-temporal feature matrices E1 and E2 of the traffic points with spacial-temporal information; the adjacency matrix of the i-th view hypergraph is calculated according to the following formula

[0015]

[0016]

[0017]

[0018] In the above formula, p i represents the size of the i-th view, and Idx represents the index number or position of the p i largest element in each row of the matrix H; in the finally obtained adjacency matrix of the i-th view hypergraph , the same position elements as the p i largest elements in each row of the matrix H are set to 1, and other position elements are set to 0; the matrix H is calculated by a multi-layer perception MLP network comprising tanh units and ReLU units;

[0019] (23) calculating all n hypergraphs representing the correlation between traffic points according to step (22), and each hypergraph has a different view size.

[0020] The time convolution function submodule of the first spatio-temporal feature processing layer of the spatio-temporal feature processing module receives the high-dimensional historical traffic flow data output by the input convolution module as input, and the time convolution function submodules of the remaining layers receive the output of the correlation modeling function submodule of the previous layer as input;

[0021] The time convolution function submodule adopts an expanded perception convolution layer to extract different scale time features of the traffic flow data; the expanded perception convolution layer comprises four convolution kernels with different sizes, and the sizes are 1x2, 1x3, 1x6 and 1x7 respectively; under different sampling granularities, the convolution kernels with different sizes can capture the time patterns of traffic flow indicators with different periods; the specific calculation process is as follows:

[0022] Given traffic flow data with a length of T and a convolution kernel f 1×k with a size of k, the convolution operation is defined as:

[0023]

[0024] In the above formula, ★ is a convolution operator, and d is an expansion factor;

[0025] After extraction through n different size convolution kernels, n high-dimensional features of multi-scale time patterns are obtained, the n high-dimensional features of multi-scale time patterns are spliced to obtain high-dimensional features o' after splicing, and then o' is input to a gating mechanism to obtain the final output of the time convolution function sub-module, i.e. high-dimensional features fused with space-time information; wherein the gating mechanism is executed by sigmoid and tanh functions, and the specific process is as follows:

[0026]

[0027] o = sigmoid(o') o tanh(o')

[0028] In each space-time feature processing layer, the time convolution sub-module performs the above feature extraction process on each traffic point to obtain a traffic point high-dimensional feature matrix of the layer, and inputs the traffic point high-dimensional feature matrix to the correlation modeling function sub-module and the output convolution module of the same layer.

[0029] The correlation modeling function sub-module performs multi-view feature fusion on each traffic point according to the n hypergraphs output by the correlation capture function module and the traffic point high-dimensional feature matrix output by the time convolution function sub-module of the layer; the specific process includes two parts of correlation modeling and multi-view feature fusion.

[0030] The correlation modeling part includes two iteratively performed stages of "node-hyperedge" and "hyperedge-node", which are used to model the correlation between traffic points in detail and aggregate the spatial features of specific points; the multi-view feature fusion part is used to comprehensively consider multi-view information to obtain updated traffic point spatial features.

[0031] The correlation modeling part performs fine-grained calculation on the potential types of the correlation between traffic points in the "node-hyperedge" stage, and distinguishes different types of correlation to aggregate the features between related nodes in the "hyperedge-node" stage, to accurately calculate the spatial features of traffic point nodes;

[0032] The specific calculation process is as follows:

[0033] In the "node-hyperedge" stage, given the traffic point high-dimensional feature matrix output by the time convolution function sub-module of the same layer This stage obtains the high-dimensional feature matrix of the hyperedge by performing hypergraph convolution operation Where D n and D e are the feature dimensions of the nodes and hyperedges respectively, and the hypergraph convolution operation is as follows:

[0034]

[0035] where Θ n are learnable parameters, and then the type of hyperedge is calculated based on Gumble Softmax operation:

[0036]

[0037] where is the hyperedge type probability matrix, where M and K represent the number of hyperedges and the number of all potential types respectively, g is an independent and identically distributed sample from Gumble (0, 1) distribution, τ is a temperature coefficient for controlling the smoothness of sampling, and element c m,k represents the probability of the mth hyperedge belonging to the kth association type; through the inference of the potential association type, the fine-grained modeling of the multi-element association type between the traffic points is realized, so as to distinguish the different types of influence relationship between the traffic points;

[0038] According to the inferred association type, the attributes of the hyperedge are further calculated as follows:

[0039]

[0040] where indicates the probability of each hyperedge belonging to the kth association type, that is, the kth column of the hyperedge type probability matrix, and is the Hadamard product representing element-wise multiplication of matrices; MLP k is a multilayer perceptron for processing the kth association type hyperedge;

[0041] In the "hyperedge-node" stage, the hypergraph convolution operation is performed to aggregate the features of the hyperedges related to the node, and the high-dimensional spatiotemporal features of the traffic point are obtained:

[0042]

[0043] where Θ e are learnable parameters;

[0044] The above calculation process is performed on all n views of the hypergraph, and finally n single-view high-dimensional spatiotemporal features of each traffic point are obtained.

[0045] The calculation process of the multi-view feature fusion part is as follows: first, sequentially connect the n single-view high-dimensional spatiotemporal features of each traffic point, then perform dilated convolution operation with a convolution kernel of size 1x n, and the dilation coefficient is the dimension of the single-view high-dimensional spatiotemporal feature.

[0046] The application also proposes a traffic flow prediction method based on fine association mining, which includes the following operation steps:

[0047] (71)The input convolution module maps the low-dimensional historical traffic flow data perceived by the road traffic point sensor into high-dimensional historical traffic flow data, and then inputs the high-dimensional historical traffic flow data into the space-time processing module;

[0048] (72)The correlation capture function module adaptively extracts the correlation between traffic points, and calculates n hypergraphs representing the correlation between traffic points from n different perspectives;

[0049] (73)The space-time feature processing module receives the high-dimensional historical traffic flow data output by the input convolution module and the n hypergraphs output by the correlation capture function module, calculates l high-dimensional traffic point feature matrices of each traffic point fused with space-time information, and then outputs to the output convolution module;

[0050] (74)The output convolution module directly sums the l high-dimensional traffic point feature matrices of each traffic point output by the space-time feature processing module, and then maps the summed feature matrices to a low-dimensional traffic flow data matrix, i.e. the predicted value of the traffic point flow speed in the future set time period, using a 1x1 standard convolution layer.

[0051] The system and method proposed by the application are beneficial in that the system and method proposed by the application are based on the correlation capture function of the multi-perspective hypergraph, which adaptively discovers complex correlation characteristics from data in view of the complex correlation characteristics of road network point traffic flow data under different scales, extracts multi-scale time features, describes potential types of correlation between road network points, aggregates spatial features of data, and realizes accurate traffic flow prediction in view of the inherent complex space-time correlation of traffic data. BRIEF DESCRIPTION OF DRAWINGS

[0052] Figure 1 is a module schematic diagram of the traffic flow prediction system based on fine correlation mining proposed by the application;

[0053] Figure 2 is a structure schematic diagram of the space-time feature processing layer proposed by the application;

[0054] Figure 3 is a calculation process schematic diagram of the output convolution module in the embodiment of the application;

[0055] Figure 4 is a calculation process schematic diagram of the correlation capture function module in the embodiment of the application;

[0056] Figure 5 is a calculation process schematic diagram of one hypergraph example in the embodiment of the application;

[0057] Figure 6 is a calculation process schematic diagram of the time convolution function sub-module in the embodiment of the application;

[0058] Figure 7 is an example diagram of different size convolution kernels capturing different periodic traffic flows in embodiments of the present application;

[0059] Figure 8 is a schematic diagram of the calculation process of the correlation modeling function sub-module in embodiments of the present application;

[0060] Figure 9 is a flowchart of the traffic prediction method based on fine correlation mining proposed in the present application; DETAILED DESCRIPTION

[0061] To make the purpose, technical solutions and advantages of the present application clearer, the present application is described in further detail below with reference to the drawings.

[0062] Referring to Figure 1 , a traffic prediction system based on fine correlation mining proposed in the present application is introduced, which includes the following modules:

[0063] The input convolution module: the function of this module is to map the low-dimensional historical traffic flow data sensed by the sensors deployed at the road traffic points to high-dimensional historical traffic flow data, and then input the high-dimensional historical traffic flow data to the space-time processing module; the input convolution module is composed of a standard convolution network layer with a kernel size of 1x1;

[0064] The correlation capture function module: the function of this module is to adaptively extract the correlation between traffic points, that is, from n different perspectives, n hypergraphs representing the correlation between traffic points are calculated, where n is a natural number, and in the embodiment, the value is 2; the nodes in the hypergraph correspond to traffic points; each hyperedge connects several nodes, representing the correlation shared between the connected traffic points, such as similar morning and evening peak patterns, a tendency to appear congestion at the same time, traffic points prone to accidents, etc.; the hypergraph is represented by an adjacency matrix , where subscript i represents the hypergraph of the i-th perspective, p i represents the size of the i-th perspective, and N represents the number of traffic points deployed on the road; E i represents the set of hyperedges in the i-th perspective hypergraph; the correlation capture function module is composed of a multi-layer perceptron (MLP) network connection containing tanh units and ReLU units;

[0065] The space-time feature processing module: this module is composed of l network structures of the same space-time feature processing layer, where l is a natural number, and in the embodiment, the value is 3. Referring to Figure 2each spatio-temporal feature processing layer is connected by a time convolution function sub-module and an association modeling function sub-module; the function of the module is to receive the high-dimensional historical traffic flow data output by the input convolution module and the n supergraphs output by the association capturing function module, calculate the l high-dimensional feature matrices of each traffic point position fused with spatio-temporal information, and then output to the output convolution module;

[0066] Referring to Figure 3 , the output convolution module: the function of the module is to directly sum the l high-dimensional feature matrices of each traffic point position output by the spatio-temporal feature processing module, and then map the high-dimensional feature matrices summed before to a low-dimensional traffic flow data matrix, i.e. the predicted value of the traffic point position flow speed in the future set time period, by using a 1x1 standard convolution layer;

[0067] Referring to Figure 4 The specific calculation process of the association capturing function module to calculate n supergraphs representing the association relationship between traffic points includes the following operation steps:

[0068] (21) Given historical traffic flow data where T p represents the length of the historical period, and N represents the number of traffic points deployed on the road; the association capturing function module first randomly initializes a spatial feature matrix for the N traffic points where D1 is the set spatial feature dimension, and then generates a time feature for the N traffic points according to the timestamp information of the collected data, i.e. obtains a time feature matrix from the given historical traffic flow data X, where D2 is the set time feature dimension; by concatenating the spatial feature matrix and the time feature matrix in the feature dimension, and using a multi-layer perception MLP to encode the concatenated matrix, a spatio-temporal feature matrix of the traffic point position fused with spatio-temporal information is obtained where each row of the spatio-temporal feature matrix E represents the spatio-temporal feature vector of the corresponding numbered traffic point position;

[0069] (22) According to step (21), two spatial feature matrices E s1 and E s2 are randomly generated, which are concatenated with the time feature matrix E t respectively, and then encoded by a multi-layer perception to obtain two spatio-temporal feature matrices E1 and E2 of the traffic point position fused with spatio-temporal information; the adjacency matrix of the i-th perspective supergraph is calculated according to the following formula

[0070]

[0071]

[0072]

[0073] In the above formula, p i Idx represents the size of the i-th viewpoint, and Idx represents the largest value p in each row of matrix H. i The index number of each element represents its position; the adjacency matrix of the i-th view hypergraph is obtained at the end. In the matrix H, p is the largest value in each row. i Elements at the same position are set to 1, and elements at other positions are set to 0; the matrix H is calculated by a multilayer perceptron (MLP) network containing tanh units and ReLU units;

[0074] (23) Calculate all n hypergraphs representing the relationships between traffic points according to step (22), and each hypergraph has a different view size.

[0075] See Figure 5 The above calculation process is illustrated with an example as follows: Assume there are 7 traffic points in the road network, and set the viewing angle to 3, i.e., p i =3, given historical traffic flow data, road network structure information, and viewpoint size, the association capture function module adaptively learns the adjacency matrix of the hypergraph. This matrix reflects the similarity between high-dimensional features of traffic points in the road network. This module assigns the most relevant p to each traffic point based on similarity ranking. i -1 = 2 other points, connected by hyperedges, forming a group of traffic points with shared association relationships.

[0076] The temporal convolution function submodule of the first spatiotemporal feature processing layer of the spatiotemporal feature processing module receives the high-dimensional historical traffic flow data output by the input convolution module as input, and the temporal convolution function submodules of the remaining layers receive the output of the correlation modeling function submodule of the layer above it as input.

[0077] See Figure 6 The temporal convolution function submodule employs a dilated perceptual convolutional layer to extract temporal features of traffic flow data at different scales; the dilated perceptual convolutional layer contains four convolutional kernels of different sizes, namely 1×2, 1×3, 1×6, and 1×7; see [link to documentation]. Figure 7 At different sampling granularities, convolutional kernels of different sizes can capture the temporal patterns of traffic flow indicators at different periods. For example, given a sampling granularity of 5 minutes, a 1×2 convolutional kernel and a 1×6 convolutional kernel can capture the temporal variation patterns of traffic flow data within the field of view for 10 minutes and 30 minutes, respectively.

[0078] The specific calculation process is as follows:

[0079] Given traffic flow data of length T and a convolution kernel f of size k 1×k The convolution operation is defined as:

[0080]

[0081] In the above formula, * is a convolution operator, d is an inflation factor, and in an embodiment, d is 1;

[0082] After extraction of n different size convolution kernels, n high-dimensional features of multi-scale time patterns are obtained, the n high-dimensional features of multi-scale time patterns are spliced to obtain a spliced high-dimensional feature o', and then o' is input to a gating mechanism to obtain the final output of the time convolution function sub-module, i.e., a high-dimensional feature fused with space-time information; wherein the gating mechanism is executed by a sigmoid and a tanh function, and the specific process is as follows:

[0083]

[0084] o = sigmoid(o') o tanh(o')

[0085] In each space-time feature processing layer, the time convolution sub-module performs the above feature extraction process on each traffic point to obtain a traffic point high-dimensional feature matrix of the layer, and inputs the traffic point high-dimensional feature matrix to the correlation modeling function sub-module and the output convolution module of the same layer.

[0086] Referring to Figure 8 The correlation modeling function sub-module performs multi-view feature fusion on each traffic point according to the n hypergraphs output by the correlation capturing function module and the traffic point high-dimensional feature matrix output by the time convolution function sub-module of the layer; the specific process includes two parts of correlation modeling and multi-view feature fusion.

[0087] The correlation modeling part includes two iteratively performed stages of "node-hyperedge" and "hyperedge-node", which are used to model the correlation between traffic points in detail and aggregate the spatial features of specific points; the multi-view feature fusion part is used to comprehensively consider multi-view information to obtain updated traffic point spatial features.

[0088] The correlation modeling part performs fine-grained calculation on the potential types of the correlation between traffic points in the "node-hyperedge" stage, and distinguishes different types of correlation to aggregate the features between the related nodes in the "hyperedge-node" stage, to accurately calculate the spatial features of the traffic point nodes;

[0089] The specific calculation process is as follows:

[0090] In the "node-superedge" stage, the traffic point high-dimensional feature matrix output by the same layer time convolution function submodule is given The high-dimensional feature matrix of the superedge is obtained by performing a supergraph convolution operation in this stage Where D n and D e are the feature dimensions of the nodes and superedges respectively, and the supergraph convolution operation is as follows:

[0091]

[0092] Where Θ n is a learnable parameter, and then the type of the superedge is calculated based on the Gumble Softmax operation:

[0093]

[0094] Where is the superedge type probability matrix, where M and K represent the number of superedges and the number of all potential types respectively, g is an independent and identically distributed sample obtained from the Gumble(0,1) distribution, τ is a temperature coefficient for controlling the sampling smoothness, and in the embodiment, the value is 0.5, and the element c m,k represents the probability that the mth superedge belongs to the kth association type; by inferring the potential association type, the fine-grained modeling of the multi-element association type between traffic points is realized, and the influence relationship between traffic points of different types is distinguished; according to the inferred association type, the attributes of the superedge are further calculated as follows:

[0095]

[0096] Where indicates the probability that each superedge belongs to the kth association type, i.e., the kth column of the superedge type probability matrix, and is the Hadamard product representing element-wise multiplication of matrices; MLP k is a multilayer perceptron for processing the kth association type superedge;

[0097] In the "superedge-node" stage, the supergraph convolution operation is performed to aggregate the features of the superedges related to the nodes, and the high-dimensional spatiotemporal features of the traffic points are obtained:

[0098]

[0099] Where Θ e is a learnable parameter;

[0100] The above calculation process is performed on all n-view supergraphs, and finally the n single-view high-dimensional spatiotemporal features of each traffic point are obtained.

[0101] The calculation process of the multi-view feature fusion part is as follows: first, sequentially connecting n single-view high-dimensional spatio-temporal features of each traffic point, then performing dilated convolution operation with a convolution kernel with a size of 1x n, and the dilated coefficient is the dimension of the single-view high-dimensional spatio-temporal feature.

[0102] Referring to Figure 9 The application also provides a traffic flow prediction method based on fine association mining, which comprises the following operation steps:

[0103] (71) The input convolution module maps the low-dimensional historical traffic flow data sensed by the road traffic point sensor into high-dimensional historical traffic flow data, and then inputs the high-dimensional historical traffic flow data into the spatio-temporal processing module;

[0104] (72) The association capturing function module adaptively extracts the association relationship between traffic points, and calculates n hypergraphs representing the association relationship between traffic points from n different views;

[0105] (73) The spatio-temporal feature processing module receives the high-dimensional historical traffic flow data output by the input convolution module and the n hypergraphs output by the association capturing function module, calculates l traffic point high-dimensional feature matrices of each traffic point fused with spatio-temporal information, and then outputs to the output convolution module;

[0106] (74) The output convolution module directly sums the l traffic point high-dimensional feature matrices of each traffic point output by the spatio-temporal feature processing module, and then uses a standard convolution layer with a size of 1x1 to map the summed feature matrix into a low-dimensional traffic flow data matrix, i.e. the predicted value of the traffic point flow speed in the future set time period.

[0107] The inventors conducted experiments on two traffic flow prediction public data sets PEMS-BAY and PEMS-D8. The basic information of the data sets is shown in Table 1:

[0108] Table 1

[0109]

[0110] For each traffic point, the task is to predict the traffic flow data of the future 12 sampling points with the historical 12 sampling points, i.e. to predict the future 1 hour flow speed with the historical 1 hour flow speed. The results are measured by three commonly used evaluation indexes of prediction task, i.e. mean absolute error (MAE), root mean square error (RMSE) and mean absolute percentage error (MAPE). Given the true value y and the predicted value The evaluation index calculation method is as follows:

[0111]

[0112]

[0113]

[0114] where y i and respectively represent the ith real value and predicted value, the smaller the three evaluation indexes, the better the prediction performance. The experimental results on each data set are shown in Table 2:

[0115] Table 2

[0116]

[0117] In the experiment, the number of hypergraph perspectives n is set to 2, and the perspective sizes are p1=20 and p2=140 respectively. Compared with other technologies, the experimental results on the two data sets have achieved the best performance. In addition, the inventors also conducted an ablation experiment on the key modules in the proposed system on the PEMS D8 data set, and verified the role of the association capturing function module and the association modeling function sub-module in the spatio-temporal feature processing module. The experimental results are shown in Table 3:

[0118] Table 3

[0119]

[0120] In the experimental groups of Table 3, the specific experiment of “w / o association capturing function module” is to only capture the single perspective association hypergraph matrix, wherein the perspective size p1=20 is set; the specific experiment of “w / o association modeling function sub-module” is to set the number of association types to 1, that is, only pay attention to whether there is an association between the traffic points, without modeling the potential association types. The ablation experiment results verify the effectiveness of the key modules in the proposed system.

[0121] The experimental results show that the system and method proposed by the present application are effective and feasible.

Claims

1. A traffic prediction system based on fine-grained correlation mining, characterized in that: The system includes the following modules: Input Convolution Module: The function of this module is to map the low-dimensional historical traffic flow data sensed by sensors deployed at road traffic points into high-dimensional historical traffic flow data, and then input the high-dimensional historical traffic flow data into the spatiotemporal processing module; the input convolution module consists of a standard convolutional network layer with a convolution kernel size of 1×1; Association Capture Function Module: This module adaptively extracts the associations between traffic points. Specifically, it calculates n hypergraphs representing the associations between traffic points from n different perspectives. Nodes in these hypergraphs correspond to traffic points. Each hyperedge connects several nodes, representing the shared associations among these connected traffic points. The hypergraphs use adjacency matrices. Let p represent the hypergraph, where the subscript i indicates the hypergraph from the i-th viewpoint. i E represents the size of the i-th viewpoint, and N represents the number of traffic points deployed on the road; i The set of hyperedges in the hypergraph representing the i-th viewpoint; the association capture function module is composed of a multilayer perceptron (MLP) network containing tanh units and ReLU units; Spatiotemporal feature processing module: This module consists of l spatiotemporal feature processing layers with the same network structure. Each spatiotemporal feature processing layer is composed of a temporal convolution function submodule and an association modeling function submodule connected together. The function of this module is to receive the high-dimensional historical traffic flow data output by the input convolution module and the n hypergraphs output by the association capture function module, calculate l high-dimensional feature matrices that fuse spatiotemporal information for each traffic point, and then output them to the output convolution module. Output Convolution Module: The function of this module is to directly sum the l high-dimensional feature matrices of each traffic point output by the spatiotemporal feature processing module, and then use a 1×1 standard convolutional layer to map the summed high-dimensional feature matrix into a low-dimensional traffic flow data matrix, that is, the predicted value of the traffic flow speed of the traffic point within a set time period in the future.

2. The traffic prediction system based on fine-grained correlation mining according to claim 1, characterized in that: The process by which the association capture function module calculates n hypergraphs representing the relationships between traffic points from n different perspectives includes the following steps: (21) Given historical traffic flow data Where T p The length of the historical period is represented by N, and the number of traffic points deployed on the road is represented by N. The associated capture function module first randomly initializes a spatial feature matrix for the N traffic points. Where D1 is the defined spatial feature dimension, and then time features are generated for N traffic points based on the timestamp information of the collected data, that is, a time feature matrix is ​​obtained from the given historical traffic flow data X. Where D2 is the defined temporal feature dimension; by concatenating the aforementioned spatial feature matrix and temporal feature matrix along the feature dimension, and using a multilayer perceptron (MLP) to encode the concatenated matrix, a spatiotemporal feature matrix of traffic points integrating spatiotemporal information is obtained. Each row of the spatiotemporal feature matrix E represents the spatiotemporal feature vector of the corresponding traffic point. (22) Following step (21), randomly generate two spatial feature matrices E s1 and E s2 , respectively with the time feature matrix E t The two matrices are concatenated, and then encoded separately using a multilayer perceptron to obtain two spatiotemporal feature matrices E1 and E2 of traffic points with fused spatiotemporal information. The adjacency matrix of the i-th view hypergraph is calculated according to the following formula. In the above formula, p i Idx represents the size of the i-th viewpoint, and Idx represents the largest value p in each row of matrix H. i The index number of each element represents its position; the adjacency matrix of the i-th view hypergraph is obtained at the end. In the matrix H, p is the largest value in each row. i Elements at the same position are set to 1, and elements at other positions are set to 0; the matrix H is calculated by a multilayer perceptron (MLP) network containing tanh units and ReLU units; (23) Calculate all n hypergraphs representing the relationships between traffic points according to step (22), and each hypergraph has a different view size.

3. The traffic prediction system based on fine-grained correlation mining according to claim 1, characterized in that: The temporal convolution function submodule of the first spatiotemporal feature processing layer of the spatiotemporal feature processing module receives the high-dimensional historical traffic flow data output by the input convolution module as input, and the temporal convolution function submodules of the remaining layers receive the output of the correlation modeling function submodule of the layer above it as input. The temporal convolution submodule employs a dilated awareness convolutional layer to extract temporal features from traffic flow data at different scales. This dilated awareness convolutional layer contains four convolutional kernels of different sizes: 1×2, 1×3, 1×6, and 1×7. At different sampling granularities, convolutional kernels of different sizes can capture temporal patterns of traffic flow indicators at different periods. The specific calculation process is as follows: Given traffic flow data of length T and a convolution kernel f of size k 1×k The convolution operation is defined as: In the above formula, ★ represents the convolution operator, and d is the dilation factor; After extraction by n convolutional kernels of different sizes, n high-dimensional features of the multi-scale temporal pattern are obtained. These n high-dimensional features are then concatenated to obtain the concatenated high-dimensional feature o′. o′ is then input into a gating mechanism to obtain the final output of the temporal convolutional function submodule, which is a high-dimensional feature that integrates spatiotemporal information. The gating mechanism is executed by the sigmoid and tanh functions, and the specific process is as follows: o=sigmoid(o′)⊙tanh(o′) In each of the spatiotemporal feature processing layers, the temporal convolution submodule performs the above-described feature extraction process on each traffic point to obtain a high-dimensional feature matrix of the traffic points in that layer, and inputs the high-dimensional feature matrix of the traffic points to the association modeling function submodule and the output convolution module in the same layer.

4. The traffic prediction system based on fine-grained correlation mining according to claim 1 or 3, characterized in that: The association modeling function submodule performs multi-view feature fusion on each traffic point based on the n hypergraphs output by the association capture function module and the high-dimensional feature matrix of traffic points output by the temporal convolution function submodule of this layer. The specific process includes two parts: correlation modeling and multi-view feature fusion. The association modeling part includes two iterative stages: "node-hyperedge" and "hyperedge-node", which are used to model the association between traffic points in a fine-grained manner and aggregate the spatial features of specific points; the multi-view feature fusion part is used to comprehensively consider multi-view information to obtain updated spatial features of traffic points.

5. The traffic prediction system based on fine-grained correlation mining according to claim 4, characterized in that: The aforementioned association modeling part performs fine-grained calculations on the potential types of associations between traffic points in the "node-hyperedge" stage, and distinguishes different types of associations in the "hyperedge-node" stage to aggregate the features between related nodes and accurately calculate the spatial features of traffic point nodes. The specific calculation process is as follows: In the "node-hyperedge" stage, given the high-dimensional feature matrix of traffic points output by the aforementioned temporal convolutional function submodule, This stage obtains the high-dimensional feature matrix of the hyperedge by performing hypergraph convolution operations. Where D n and D e Given the defined feature dimensions of the nodes and hyperedges, respectively, the hypergraph convolution operation is as follows: Where Θ n The parameters are then learned, and the type of the hyperedge is calculated based on the Gumble Softmax operation: in Let be the hyperedge type probability matrix, where M and K represent the number of hyperedges and the total number of all potential types, respectively; g is the independent and identically distributed sampled from the Gumble(0,1) distribution; τ is the temperature coefficient controlling the sampling smoothness; and c is the element. m,k This represents the probability that the m-th hyperedge belongs to the k-th association type. By inferring the potential association types, fine-grained modeling of multiple association types between traffic points is achieved, enabling the differentiation of different types of influence relationships between traffic points. Based on the inferred association type, the attributes of the hyperedge are further calculated as follows: in This indicates the probability that each hyperedge belongs to the k-th association type, i.e., the k-th column of the hyperedge type probability matrix. ⊙ represents the Hadamard product, indicating element-wise matrix multiplication. MLP k A multilayer perceptron for processing the k-th associated type hyperedge; In the "hyperedge-node" stage, hypergraph convolution operations are performed to aggregate the features of hyperedges associated with nodes, resulting in high-dimensional spatiotemporal features of traffic points: Where Θ e These are learnable parameters; The above calculation process is performed on all n viewpoints of the hypergraph, and finally, n single-view high-dimensional spatiotemporal features of each traffic point are obtained.

6. The traffic prediction system based on fine-grained correlation mining according to claim 5, characterized in that: The calculation process of the multi-view feature fusion part is as follows: First, the n single-view high-dimensional spatiotemporal features of each traffic point are sequentially connected, and then dilated convolution operation is performed using a convolution kernel of size 1×n, with the dilation coefficient being the dimension of the single-view high-dimensional spatiotemporal feature.

7. A traffic prediction method based on refined correlation mining, characterized in that: The method includes the following steps: (71) The input convolution module maps the low-dimensional historical traffic flow data sensed by the sensors deployed at road traffic points into high-dimensional historical traffic flow data, and then inputs the high-dimensional historical traffic flow data into the spatiotemporal processing module. (72) The association capture function module adaptively extracts the association between traffic points and calculates n hypergraphs representing the association between traffic points from n different perspectives. (73) The spatiotemporal feature processing module receives the high-dimensional historical traffic flow data output by the input convolution module and the n hypergraphs output by the association capture function module, calculates the l high-dimensional feature matrix of each traffic point that integrates spatiotemporal information, and then outputs it to the output convolution module. (74) The output convolution module directly sums the high-dimensional feature matrices of each traffic point output by the spatiotemporal feature processing module, and then uses a 1×1 standard convolutional layer to map the summed feature matrix into a low-dimensional traffic flow data matrix, that is, the predicted value of the traffic flow speed of the traffic point within a set time period in the future.

Images